[
{
"path": ".clang-format",
"content": "# This file is used by clang-format to autoformat paddle source code\n#\n# The clang-format is part of llvm toolchain.\n# It need to install llvm and clang to format source code style.\n#\n# The basic usage is,\n# clang-format -i -style=file PATH/TO/SOURCE/CODE\n#\n# The -style=file implicit use \".clang-format\" file located in one of\n# parent directory.\n# The -i means inplace change.\n#\n# The document of clang-format is\n# http://clang.llvm.org/docs/ClangFormat.html\n# http://clang.llvm.org/docs/ClangFormatStyleOptions.html\n---\nLanguage: Cpp\nBasedOnStyle: Google\nIndentWidth: 4\nTabWidth: 4\nContinuationIndentWidth: 4\nMaxEmptyLinesToKeep: 2\nAccessModifierOffset: -2 # The private/protected/public has no indent in class\nStandard: Cpp11\nAllowAllParametersOfDeclarationOnNextLine: true\nBinPackParameters: false\nBinPackArguments: false\n...\n\n"
},
{
"path": ".flake8",
"content": "[flake8]\n\n########## OPTIONS ##########\n# Set the maximum length that any line (with some exceptions) may be.\nmax-line-length = 120\n\n\n################### FILE PATTERNS ##########################\n# Provide a comma-separated list of glob patterns to exclude from checks.\nexclude =\n # git folder\n .git,\n # python cache\n __pycache__,\n # third party\n utils/compute-wer.py,\n third_party/,\n# Provide a comma-separate list of glob patterns to include for checks.\nfilename =\n *.py\n\n\n########## RULES ##########\n\n# ERROR CODES\n#\n# E/W - PEP8 errors/warnings (pycodestyle)\n# F - linting errors (pyflakes)\n# C - McCabe complexity error (mccabe)\n#\n# W503 - line break before binary operator\n\n# Specify a list of codes to ignore.\nignore =\n W503\n E252,E262,E127,E265,E126,E266,E241,E261,E128,E125,E129\n W291,W293,W605\n E203,E305,E402,E501,E721,E741,F403,F405,F821,F841,F999,W503,W504,C408,E302,W291,E303,\n # shebang has extra meaning in fbcode lints, so I think it's not worth trying\n # to line this up with executable bit\n EXE001,\n # these ignores are from flake8-bugbear; please fix!\n B007,B008,\n # these ignores are from flake8-comprehensions; please fix!\n C400,C401,C402,C403,C404,C405,C407,C411,C413,C414,C415\n\n\nper-file-ignores =\n */__init__.py: F401\n\n# Specify the list of error codes you wish Flake8 to report.\nselect =\n E,\n W,\n F,\n C\n"
},
{
"path": ".gitconfig",
"content": "[alias]\n st = status\n ci = commit\n br = branch\n co = checkout\n df = diff\n l = log --pretty=format:\\\"%h %ad | %s%d [%an]\\\" --graph --date=short\n ll = log --stat\n\n[merge]\n tool = vimdiff\n\n[core]\n excludesfile = ~/.gitignore\n editor = vim\n\n[color]\n branch = auto\n diff = auto\n status = auto\n\n[color \"branch\"]\n current = yellow reverse\n local = yellow\n remote = green\n\n[color \"diff\"]\n meta = yellow bold\n frag = magenta bold\n old = red bold\n new = green bold\n\n[color \"status\"]\n added = yellow\n changed = green\n untracked = cyan\n\n[push]\n default = matching\n\n[credential]\n helper = store\n\n[user]\n name =\n email =\n\n\n"
},
{
"path": ".github/CODE_OF_CONDUCT.md",
"content": "# Contributor Covenant Code of Conduct\n\n## Our Pledge\n\nIn the interest of fostering an open and welcoming environment, we as\ncontributors and maintainers pledge to making participation in our project and\nour community a harassment-free experience for everyone, regardless of age, body\nsize, disability, ethnicity, sex characteristics, gender identity and expression,\nlevel of experience, education, socio-economic status, nationality, personal\nappearance, race, religion, or sexual identity and orientation.\n\n## Our Standards\n\nExamples of behavior that contributes to creating a positive environment\ninclude:\n\n* Using welcoming and inclusive language\n* Being respectful of differing viewpoints and experiences\n* Gracefully accepting constructive criticism\n* Focusing on what is best for the community\n* Showing empathy towards other community members\n\nExamples of unacceptable behavior by participants include:\n\n* The use of sexualized language or imagery and unwelcome sexual attention or\n advances\n* Racial or political allusions \n* Trolling, insulting/derogatory comments, and personal or political attacks\n* Public or private harassment\n* Publishing others' private information, such as a physical or electronic\n address, without explicit permission\n* Other conduct which could reasonably be considered inappropriate in a\n professional setting\n\n## Our Responsibilities\n\nProject maintainers are responsible for clarifying the standards of acceptable\nbehavior and are expected to take appropriate and fair corrective action in\nresponse to any instances of unacceptable behavior.\n\nProject maintainers have the right and responsibility to remove, edit, or\nreject comments, commits, code, wiki edits, issues, and other contributions\nthat are not aligned to this Code of Conduct, or to ban temporarily or\npermanently any contributor for other behaviors that they deem inappropriate,\nthreatening, offensive, or harmful.\n\n## Scope\n\nThis Code of Conduct applies both within project spaces and in public spaces\nwhen an individual is representing the project or its community. Examples of\nrepresenting a project or community include using an official project e-mail\naddress, posting via an official social media account, or acting as an appointed\nrepresentative at an online or offline event. Representation of a project may be\nfurther defined and clarified by project maintainers.\n\n## Enforcement\n\nInstances of abusive, harassing, or otherwise unacceptable behavior may be\nreported by contacting the project team at paddlespeech@baidu.com. All\ncomplaints will be reviewed and investigated and will result in a response that\nis deemed necessary and appropriate to the circumstances. The project team is\nobligated to maintain confidentiality with regard to the reporter of an incident.\nFurther details of specific enforcement policies may be posted separately.\n\nProject maintainers who do not follow or enforce the Code of Conduct in good\nfaith may face temporary or permanent repercussions as determined by other\nmembers of the project's leadership.\n\n## Attribution\n\nThis Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,\navailable at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html\n\n[homepage]: https://www.contributor-covenant.org\n\nFor answers to common questions about this code of conduct, see\nhttps://www.contributor-covenant.org/faq\n"
},
{
"path": ".github/CONTRIBUTING.md",
"content": "# 💡 paddlespeech 提交代码须知\n\n### Discussed in https://github.com/PaddlePaddle/PaddleSpeech/discussions/1326\n\n\n\nOriginally posted by **yt605155624** January 12, 2022\n1. 写完代码之后可以用我们的 pre-commit 检查一下代码格式，注意只改自己修改的代码的格式即可，其他的代码有可能也被改了格式，不要 add 就好\n```\npip install pre-commit\npre-commit run --file 你修改的代码\n```\n2. 提交 commit 中增加必要信息跳过不必要的 CI\n- 提交 asr 相关代码\n```text\ngit commit -m \"xxxxxx, test=asr\"\n```\n- 提交 tts 相关代码\n```text\ngit commit -m \"xxxxxx, test=tts\"\n```\n- 仅修改文档\n```text\ngit commit -m \"xxxxxx, test=doc\"\n```\n注意：\n1. 虽然跳过了 CI，但是还要先排队排到才能跳过，所以非自己方向看到 pending 不要着急 🤣\n2. 在 `git commit --amend` 的时候才加 `test=xxx` 可能不太有效\n3. 一个 pr 多次提交 commit 注意每次都要加 `test=xxx`，因为每个 commit 都会触发 CI\n4. 删除 python 环境中已经安装好的 paddlespeech，否则可能会影响 import paddlespeech 的顺序
\n"
},
{
"path": ".github/ISSUE_TEMPLATE/bug-report-s2t.md",
"content": "---\nname: \"\\U0001F41B S2T Bug Report\"\nabout: Create a report to help us improve\ntitle: \"[S2T]XXXX\"\nlabels: Bug, S2T\nassignees: zh794390558\n\n---\n\nFor support and discussions, please use our [Discourse forums](https://github.com/PaddlePaddle/DeepSpeech/discussions).\n\nIf you've found a bug then please create an issue with the following information:\n\n**Describe the bug**\nA clear and concise description of what the bug is.\n\n**To Reproduce**\nSteps to reproduce the behavior:\n1. Go to '...'\n2. Click on '....'\n3. Scroll down to '....'\n4. See error\n\n**Expected behavior**\nA clear and concise description of what you expected to happen.\n\n**Screenshots**\nIf applicable, add screenshots to help explain your problem.\n\n**Environment (please complete the following information):**\n - OS: [e.g. Ubuntu]\n - GCC/G++ Version [e.g. 8.3]\n - Python Version [e.g. 3.7]\n - PaddlePaddle Version [e.g. 2.0.0]\n - Model Version [e.g. 2.0.0]\n - GPU/DRIVER Information [e.g. Tesla V100-SXM2-32GB/440.64.00]\n - CUDA/CUDNN Version [e.g. cuda-10.2]\n - MKL Version\n- TensorRT Version\n\n**Additional context**\nAdd any other context about the problem here.\n"
},
{
"path": ".github/ISSUE_TEMPLATE/bug-report-tts.md",
"content": "---\nname: \"\\U0001F41B TTS Bug Report\"\nabout: Create a report to help us improve\ntitle: \"[TTS]XXXX\"\nlabels: Bug, T2S\n\n---\n\nFor support and discussions, please use our [Discourse forums](https://github.com/PaddlePaddle/DeepSpeech/discussions).\n\nIf you've found a bug then please create an issue with the following information:\n\n**Describe the bug**\nA clear and concise description of what the bug is.\n\n**To Reproduce**\nSteps to reproduce the behavior:\n1. Go to '...'\n2. Click on '....'\n3. Scroll down to '....'\n4. See error\n\n**Expected behavior**\nA clear and concise description of what you expected to happen.\n\n**Screenshots**\nIf applicable, add screenshots to help explain your problem.\n\n**Environment (please complete the following information):**\n - OS: [e.g. Ubuntu]\n - GCC/G++ Version [e.g. 8.3]\n - Python Version [e.g. 3.7]\n - PaddlePaddle Version [e.g. 2.0.0]\n - Model Version [e.g. 2.0.0]\n - GPU/DRIVER Information [e.g. Tesla V100-SXM2-32GB/440.64.00]\n - CUDA/CUDNN Version [e.g. cuda-10.2]\n - MKL Version\n- TensorRT Version\n\n**Additional context**\nAdd any other context about the problem here.\n"
},
{
"path": ".github/ISSUE_TEMPLATE/feature-request.md",
"content": "---\nname: \"\\U0001F680 Feature Request\"\nabout: As a user, I want to request a New Feature on the product.\ntitle: ''\nlabels: feature request\nassignees: D-DanielYang, iftaken\n\n---\n\n## Feature Request\n\n**Is your feature request related to a problem? Please describe:**\n\n\n**Describe the feature you'd like:**\n\n\n**Describe alternatives you've considered:**\n\n"
},
{
"path": ".github/ISSUE_TEMPLATE/others.md",
"content": "---\nname: \"\\U0001F9E9 Others\"\nabout: Report any other non-support related issues.\ntitle: ''\nlabels: ''\nassignees: ''\n\n---\n\n## Others\n\n\n"
},
{
"path": ".github/ISSUE_TEMPLATE/question.md",
"content": "---\nname: \"\\U0001F914 Ask a Question\"\nabout: I want to ask a question.\ntitle: ''\nlabels: Question\nassignees: ''\n\n---\n\n## General Question\n\n\n"
},
{
"path": ".github/PULL_REQUEST_TEMPLATE.md",
"content": "\n### PR types\n\n\n### PR changes\n\n\n### Describe\n\n"
},
{
"path": ".github/stale.yml",
"content": "# Number of days of inactivity before an issue becomes stale\ndaysUntilStale: 45 \n# Number of days of inactivity before a stale issue is closed\ndaysUntilClose: 30\n# Issues with these labels will never be considered stale\nexemptLabels:\n - Roadmap \n - Bug\n - feature request\n - Tips\n# Label to use when marking an issue as stale\nstaleLabel: Stale\n# Comment to post when marking an issue as stale. Set to `false` to disable\nmarkComment: >\n This issue has been automatically marked as stale because it has not had\n recent activity. It will be closed if no further activity occurs. Thank you\n for your contributions.\nunmarkComment: false\n# Comment to post when closing a stale issue. Set to `false` to disable\ncloseComment: >\n This issue is closed. Please re-open if needed.\n"
},
{
"path": ".gitignore",
"content": ".DS_Store\n*.pyc\n.vscode\n*.log\n*.wav\n*.pdmodel\n*.pdiparams*\n*.zip\n*.tar\n*.tar.gz\n.ipynb_checkpoints\n*.npz\n*.done\n*.whl\n*.egg-info\nbuild\n*output/\n.history\n.idea\n\naudio/dist/\naudio/fc_patch/\n\ndocs/build/\ndocs/topic/ctc/warp-ctc/\n\ntools/venv\ntools/kenlm\ntools/sox-14.4.2\ntools/soxbindings\ntools/montreal-forced-aligner/\ntools/Montreal-Forced-Aligner/\ntools/sctk\ntools/sctk-20159b5/\ntools/kaldi\ntools/OpenBLAS/\ntools/Miniconda3-latest-Linux-x86_64.sh\ntools/activate_python.sh\ntools/miniconda.sh\ntools/CRF++-0.58/\ntools/liblbfgs-1.10/\ntools/srilm/\ntools/env.sh\ntools/openfst-1.8.1/\ntools/libsndfile/\ntools/python-soundfile/\ntools/onnx\ntools/onnxruntime\ntools/Paddle2ONNX\ntools/onnx-simplifier/\n\nspeechx/fc_patch/\n\nthird_party/ctc_decoders/paddlespeech_ctcdecoders.py\n\nkernel_meta/\n"
},
{
"path": ".mergify.yml",
"content": "pull_request_rules:\n - name: automatic merge for develop when CI passes and 1 reviews\n conditions:\n - \"approved-reviews-by>=1\"\n - check-success=Travis CI - Pull Request\n - base=develop\n actions:\n merge:\n method: merge\n - name: delete head branch after merged\n conditions:\n - merged\n actions:\n delete_head_branch: {}\n - name: \"add label=auto-merge for PR by mergify\"\n conditions:\n - author=mergify[bot]\n actions:\n label:\n add: [\"auto-merge\"]\n - name: warn on conflicts\n conditions:\n - conflict\n actions:\n comment:\n message: This pull request is now in conflict :(\n label:\n add: [\"conflicts\"]\n - name: unlabel conflicts\n conditions:\n - -conflict\n actions:\n label:\n remove: [\"conflicts\"]\n - name: \"auto add label=Dataset\"\n conditions:\n - files~=^dataset/\n actions:\n label:\n add: [\"Dataset\"]\n - name: \"auto add label=S2T\"\n conditions:\n - files~=^paddlespeech/s2t/\n actions:\n label:\n add: [\"S2T\"]\n - name: \"auto add label=T2S\"\n conditions:\n - files~=^paddlespeech/t2s/\n actions:\n label:\n add: [\"T2S\"]\n - name: \"auto add label=Audio\"\n conditions:\n - files~=^paddlespeech/audio/\n actions:\n label:\n add: [\"Audio\"]\n - name: \"auto add label=Vector\"\n conditions:\n - files~=^paddlespeech/vector/\n actions:\n label:\n add: [\"Vector\"]\n - name: \"auto add label=Text\"\n conditions:\n - files~=^paddlespeech/text/\n actions:\n label:\n add: [\"Text\"]\n - name: \"auto add label=Example\"\n conditions:\n - files~=^examples/\n actions:\n label:\n add: [\"Example\"]\n - name: \"auto add label=CLI\"\n conditions:\n - files~=^paddlespeech/cli\n actions:\n label:\n add: [\"CLI\"]\n - name: \"auto add label=Server\"\n conditions:\n - files~=^paddlespeech/server\n actions:\n label:\n add: [\"Server\"]\n - name: \"auto add label=Demo\"\n conditions:\n - files~=^demos/\n actions:\n label:\n add: [\"Demo\"]\n - name: \"auto add label=README\"\n conditions:\n - files~=(README.md|READEME_cn.md)\n actions:\n label:\n add: [\"README\"]\n - name: \"auto add label=Documentation\"\n conditions:\n - files~=^(docs/|CHANGELOG.md)\n actions:\n label:\n add: [\"Documentation\"]\n - name: \"auto add label=CI\"\n conditions:\n - files~=^(.circleci/|ci/|.github/|.travis.yml|.travis|env.sh)\n actions:\n label:\n add: [\"CI\"]\n - name: \"auto add label=Installation\"\n conditions:\n - files~=^(tools/|setup.py|setup.cfg|setup_audio.py)\n actions:\n label:\n add: [\"Installation\"]\n - name: \"auto add label=Test\"\n conditions:\n - files~=^(tests/)\n actions:\n label:\n add: [\"Test\"]\n - name: \"auto add label=mergify\"\n conditions:\n - files~=^.mergify.yml\n actions:\n label:\n add: [\"mergify\"]\n - name: \"auto add label=Docker\"\n conditions:\n - files~=^docker/\n actions:\n label:\n add: [\"Docker\"]\n - name: \"auto add label=Deployment\"\n conditions:\n - files~=^runtime/\n actions:\n label:\n add: [\"Deployment\"]\n"
},
{
"path": ".pre-commit-config.yaml",
"content": "repos:\n- repo: https://github.com/pre-commit/mirrors-yapf.git\n rev: v0.16.0\n hooks:\n - id: yapf\n files: \\.py$\n exclude: (?=runtime/engine/kaldi|audio/paddleaudio/src|third_party).*(\\.cpp|\\.cc|\\.h\\.hpp|\\.py)$\n\n- repo: https://github.com/pre-commit/pre-commit-hooks\n rev: a11d9314b22d8f8c7556443875b731ef05965464\n hooks:\n - id: check-merge-conflict\n - id: check-symlinks\n - id: detect-private-key\n files: (?!.*paddle)^.*$\n - id: end-of-file-fixer\n files: \\.md$\n #- id: trailing-whitespace\n # files: \\.md$\n - id: requirements-txt-fixer\n exclude: (?=third_party).*$\n - id: check-yaml\n - id: check-json\n - id: pretty-format-json\n args:\n - --no-sort-keys\n - --autofix\n - id: check-merge-conflict\n # - id: flake8\n # aergs:\n # - --ignore=E501,E228,E226,E261,E266,E128,E402,W503\n # - --builtins=G,request\n # - --jobs=1\n # exclude: (?=runtime/engine/kaldi|audio/paddleaudio/src|third_party).*(\\.cpp|\\.cc|\\.h\\.hpp|\\.py)$\n\n- repo : https://github.com/Lucas-C/pre-commit-hooks\n rev: v1.0.1\n hooks:\n - id: forbid-crlf\n files: \\.md$\n - id: remove-crlf\n files: \\.md$\n - id: forbid-tabs\n files: \\.md$\n - id: remove-tabs\n files: \\.md$\n\n- repo: local\n hooks:\n - id: clang-format\n name: clang-format\n description: Format files with ClangFormat\n entry: bash .pre-commit-hooks/clang-format.hook -i\n language: system\n files: \\.(h\\+\\+|h|hh|hxx|hpp|cuh|c|cc|cpp|cu|c\\+\\+|cxx|tpp|txx)$\n exclude: (?=runtime/engine/kaldi|audio/paddleaudio/src|runtime/patch|runtime/tools/fstbin|runtime/tools/lmbin|third_party/ctc_decoders|runtime/engine/common/utils).*(\\.cpp|\\.cc|\\.h|\\.hpp|\\.py)$ \n - id: cpplint\n name: cpplint\n description: Static code analysis of C/C++ files\n language: python\n files: \\.(h\\+\\+|h|hh|hxx|hpp|cuh|c|cc|cpp|cu|c\\+\\+|cxx|tpp|txx)$\n exclude: (?=runtime/engine/kaldi|runtime/engine/common/matrix|audio/paddleaudio/src|runtime/patch|runtime/tools/fstbin|runtime/tools/lmbin|third_party/ctc_decoders|runtime/engine/common/utils).*(\\.cpp|\\.cc|\\.h|\\.hpp|\\.py)$ \n entry: cpplint --filter=-build,-whitespace,+whitespace/comma,-whitespace/indent\n- repo: https://github.com/asottile/reorder_python_imports\n rev: v2.4.0\n hooks:\n - id: reorder-python-imports\n exclude: (?=runtime/engine/kaldi|audio/paddleaudio/src|runtime/patch|runtime/tools/fstbin|runtime/tools/lmbin|third_party/ctc_decoders).*(\\.cpp|\\.cc|\\.h\\.hpp|\\.py)$\n"
},
{
"path": ".pre-commit-hooks/clang-format.hook",
"content": "#!/usr/bin/env bash\nset -e\n\nreadonly VERSION=\"3.9\"\n\nversion=$(clang-format -version)\n\n# if ! [[ $version == *\"$VERSION\"* ]]; then\n# echo \"clang-format version check failed.\"\n# echo \"a version contains '$VERSION' is needed, but get '$version'\"\n# echo \"you can install the right version, and make an soft-link to '\\$PATH' env\"\n# exit -1\n# fi\n\nclang-format $@\n"
},
{
"path": ".pre-commit-hooks/copyright-check.hook",
"content": "# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\nimport argparse\nimport io, re\nimport sys, os\nimport subprocess\nimport platform\n\nCOPYRIGHT = '''\nCopyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n http://www.apache.org/licenses/LICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License.\n'''\n\nLANG_COMMENT_MARK = None\n\nNEW_LINE_MARK = None\n\nCOPYRIGHT_HEADER = None\n\nif platform.system() == \"Windows\":\n NEW_LINE_MARK = \"\\r\\n\"\nelse:\n NEW_LINE_MARK = '\\n'\n COPYRIGHT_HEADER = COPYRIGHT.split(NEW_LINE_MARK)[1]\n p = re.search('(\\d{4})', COPYRIGHT_HEADER).group(0)\n process = subprocess.Popen([\"date\", \"+%Y\"], stdout=subprocess.PIPE)\n date, err = process.communicate()\n date = date.decode(\"utf-8\").rstrip(\"\\n\")\n COPYRIGHT_HEADER = COPYRIGHT_HEADER.replace(p, date)\n\n\ndef generate_copyright(template, lang='C'):\n if lang == 'Python':\n LANG_COMMENT_MARK = '#'\n else:\n LANG_COMMENT_MARK = \"//\"\n\n lines = template.split(NEW_LINE_MARK)\n BLANK = \" \"\n ans = LANG_COMMENT_MARK + BLANK + COPYRIGHT_HEADER + NEW_LINE_MARK\n for lino, line in enumerate(lines):\n if lino == 0 or lino == 1 or lino == len(lines) - 1: continue\n if len(line) == 0:\n BLANK = \"\"\n else:\n BLANK = \" \"\n ans += LANG_COMMENT_MARK + BLANK + line + NEW_LINE_MARK\n\n return ans + \"\\n\"\n\n\ndef lang_type(filename):\n if filename.endswith(\".py\"):\n return \"Python\"\n elif filename.endswith(\".h\"):\n return \"C\"\n elif filename.endswith(\".c\"):\n return \"C\"\n elif filename.endswith(\".hpp\"):\n return \"C\"\n elif filename.endswith(\".cc\"):\n return \"C\"\n elif filename.endswith(\".cpp\"):\n return \"C\"\n elif filename.endswith(\".cu\"):\n return \"C\"\n elif filename.endswith(\".cuh\"):\n return \"C\"\n elif filename.endswith(\".go\"):\n return \"C\"\n elif filename.endswith(\".proto\"):\n return \"C\"\n else:\n print(\"Unsupported filetype %s\", filename)\n exit(0)\n\n\nPYTHON_ENCODE = re.compile(\"^[ \\t\\v]*#.*?coding[:=][ \\t]*([-_.a-zA-Z0-9]+)\")\n\n\ndef main(argv=None):\n parser = argparse.ArgumentParser(\n description='Checker for copyright declaration.')\n parser.add_argument('filenames', nargs='*', help='Filenames to check')\n args = parser.parse_args(argv)\n\n retv = 0\n for filename in args.filenames:\n fd = io.open(filename, encoding=\"utf-8\")\n first_line = fd.readline()\n second_line = fd.readline()\n if \"COPYRIGHT (C)\" in first_line.upper(): continue\n if first_line.startswith(\"#!\") or PYTHON_ENCODE.match(\n second_line) != None or PYTHON_ENCODE.match(first_line) != None:\n continue\n original_contents = io.open(filename, encoding=\"utf-8\").read()\n new_contents = generate_copyright(\n COPYRIGHT, lang_type(filename)) + original_contents\n print('Auto Insert Copyright Header {}'.format(filename))\n retv = 1\n with io.open(filename, 'w') as output_file:\n output_file.write(new_contents)\n\n return retv\n\n\nif __name__ == '__main__':\n exit(main())\n"
},
{
"path": ".readthedocs.yml",
"content": "# .readthedocs.yml\n# Read the Docs configuration file\n# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details\n\n# Required\nversion: 2\n\n# Build documentation in the docs/ directory with Sphinx\nsphinx:\n configuration: docs/source/conf.py\n\n# Build documentation with MkDocs\n#mkdocs:\n# configuration: mkdocs.yml\n\n# Optionally build your docs in additional formats such as PDF\nformats: []\n\n# Optionally set the version of Python and requirements required to build your docs\npython:\n version: 3.7\n install:\n - requirements: docs/requirements.txt\n - method: setuptools\n path: .\n system_packages: true\n"
},
{
"path": ".style.yapf",
"content": "[style]\nbased_on_style = pep8\ncolumn_limit = 80\n"
},
{
"path": ".travis.yml",
"content": "language: cpp\ncache: ccache\nsudo: required\ndist: Bionic \nservices:\n - docker\nos:\n - linux\nenv:\n - JOB=PRE_COMMIT\n\naddons:\n apt:\n packages:\n - git\n - python3-pip\n - python3-dev\n\nbefore_install:\n - python3 --version\n - python3 -m pip --version\n - pip3 --version\n - sudo pip3 install -U virtualenv pre-commit pip\n - docker pull paddlepaddle/paddle:latest\n\nscript:\n - exit_code=0\n - docker run -i --rm -v \"$PWD:/py_unittest\" paddlepaddle/paddle:latest /bin/bash -c\n 'cd /py_unittest && bash .travis/precommit.sh && source env.sh && bash .travis/unittest.sh' || exit_code=$(( exit_code | $? ))\n exit $exit_code\n\nnotifications:\n email:\n on_success: change\n on_failure: always\n"
},
{
"path": "LICENSE",
"content": " Apache License\n Version 2.0, January 2004\n http://www.apache.org/licenses/\n\n TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n 1. Definitions.\n\n \"License\" shall mean the terms and conditions for use, reproduction,\n and distribution as defined by Sections 1 through 9 of this document.\n\n \"Licensor\" shall mean the copyright owner or entity authorized by\n the copyright owner that is granting the License.\n\n \"Legal Entity\" shall mean the union of the acting entity and all\n other entities that control, are controlled by, or are under common\n control with that entity. For the purposes of this definition,\n \"control\" means (i) the power, direct or indirect, to cause the\n direction or management of such entity, whether by contract or\n otherwise, or (ii) ownership of fifty percent (50%) or more of the\n outstanding shares, or (iii) beneficial ownership of such entity.\n\n \"You\" (or \"Your\") shall mean an individual or Legal Entity\n exercising permissions granted by this License.\n\n \"Source\" form shall mean the preferred form for making modifications,\n including but not limited to software source code, documentation\n source, and configuration files.\n\n \"Object\" form shall mean any form resulting from mechanical\n transformation or translation of a Source form, including but\n not limited to compiled object code, generated documentation,\n and conversions to other media types.\n\n \"Work\" shall mean the work of authorship, whether in Source or\n Object form, made available under the License, as indicated by a\n copyright notice that is included in or attached to the work\n (an example is provided in the Appendix below).\n\n \"Derivative Works\" shall mean any work, whether in Source or Object\n form, that is based on (or derived from) the Work and for which the\n editorial revisions, annotations, elaborations, or other modifications\n represent, as a whole, an original work of authorship. For the purposes\n of this License, Derivative Works shall not include works that remain\n separable from, or merely link (or bind by name) to the interfaces of,\n the Work and Derivative Works thereof.\n\n \"Contribution\" shall mean any work of authorship, including\n the original version of the Work and any modifications or additions\n to that Work or Derivative Works thereof, that is intentionally\n submitted to Licensor for inclusion in the Work by the copyright owner\n or by an individual or Legal Entity authorized to submit on behalf of\n the copyright owner. For the purposes of this definition, \"submitted\"\n means any form of electronic, verbal, or written communication sent\n to the Licensor or its representatives, including but not limited to\n communication on electronic mailing lists, source code control systems,\n and issue tracking systems that are managed by, or on behalf of, the\n Licensor for the purpose of discussing and improving the Work, but\n excluding communication that is conspicuously marked or otherwise\n designated in writing by the copyright owner as \"Not a Contribution.\"\n\n \"Contributor\" shall mean Licensor and any individual or Legal Entity\n on behalf of whom a Contribution has been received by Licensor and\n subsequently incorporated within the Work.\n\n 2. Grant of Copyright License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n copyright license to reproduce, prepare Derivative Works of,\n publicly display, publicly perform, sublicense, and distribute the\n Work and such Derivative Works in Source or Object form.\n\n 3. Grant of Patent License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n (except as stated in this section) patent license to make, have made,\n use, offer to sell, sell, import, and otherwise transfer the Work,\n where such license applies only to those patent claims licensable\n by such Contributor that are necessarily infringed by their\n Contribution(s) alone or by combination of their Contribution(s)\n with the Work to which such Contribution(s) was submitted. If You\n institute patent litigation against any entity (including a\n cross-claim or counterclaim in a lawsuit) alleging that the Work\n or a Contribution incorporated within the Work constitutes direct\n or contributory patent infringement, then any patent licenses\n granted to You under this License for that Work shall terminate\n as of the date such litigation is filed.\n\n 4. Redistribution. You may reproduce and distribute copies of the\n Work or Derivative Works thereof in any medium, with or without\n modifications, and in Source or Object form, provided that You\n meet the following conditions:\n\n (a) You must give any other recipients of the Work or\n Derivative Works a copy of this License; and\n\n (b) You must cause any modified files to carry prominent notices\n stating that You changed the files; and\n\n (c) You must retain, in the Source form of any Derivative Works\n that You distribute, all copyright, patent, trademark, and\n attribution notices from the Source form of the Work,\n excluding those notices that do not pertain to any part of\n the Derivative Works; and\n\n (d) If the Work includes a \"NOTICE\" text file as part of its\n distribution, then any Derivative Works that You distribute must\n include a readable copy of the attribution notices contained\n within such NOTICE file, excluding those notices that do not\n pertain to any part of the Derivative Works, in at least one\n of the following places: within a NOTICE text file distributed\n as part of the Derivative Works; within the Source form or\n documentation, if provided along with the Derivative Works; or,\n within a display generated by the Derivative Works, if and\n wherever such third-party notices normally appear. The contents\n of the NOTICE file are for informational purposes only and\n do not modify the License. You may add Your own attribution\n notices within Derivative Works that You distribute, alongside\n or as an addendum to the NOTICE text from the Work, provided\n that such additional attribution notices cannot be construed\n as modifying the License.\n\n You may add Your own copyright statement to Your modifications and\n may provide additional or different license terms and conditions\n for use, reproduction, or distribution of Your modifications, or\n for any such Derivative Works as a whole, provided Your use,\n reproduction, and distribution of the Work otherwise complies with\n the conditions stated in this License.\n\n 5. Submission of Contributions. Unless You explicitly state otherwise,\n any Contribution intentionally submitted for inclusion in the Work\n by You to the Licensor shall be under the terms and conditions of\n this License, without any additional terms or conditions.\n Notwithstanding the above, nothing herein shall supersede or modify\n the terms of any separate license agreement you may have executed\n with Licensor regarding such Contributions.\n\n 6. Trademarks. This License does not grant permission to use the trade\n names, trademarks, service marks, or product names of the Licensor,\n except as required for reasonable and customary use in describing the\n origin of the Work and reproducing the content of the NOTICE file.\n\n 7. Disclaimer of Warranty. Unless required by applicable law or\n agreed to in writing, Licensor provides the Work (and each\n Contributor provides its Contributions) on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n implied, including, without limitation, any warranties or conditions\n of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n PARTICULAR PURPOSE. You are solely responsible for determining the\n appropriateness of using or redistributing the Work and assume any\n risks associated with Your exercise of permissions under this License.\n\n 8. Limitation of Liability. In no event and under no legal theory,\n whether in tort (including negligence), contract, or otherwise,\n unless required by applicable law (such as deliberate and grossly\n negligent acts) or agreed to in writing, shall any Contributor be\n liable to You for damages, including any direct, indirect, special,\n incidental, or consequential damages of any character arising as a\n result of this License or out of the use or inability to use the\n Work (including but not limited to damages for loss of goodwill,\n work stoppage, computer failure or malfunction, or any and all\n other commercial damages or losses), even if such Contributor\n has been advised of the possibility of such damages.\n\n 9. Accepting Warranty or Additional Liability. While redistributing\n the Work or Derivative Works thereof, You may choose to offer,\n and charge a fee for, acceptance of support, warranty, indemnity,\n or other liability obligations and/or rights consistent with this\n License. However, in accepting such obligations, You may act only\n on Your own behalf and on Your sole responsibility, not on behalf\n of any other Contributor, and only if You agree to indemnify,\n defend, and hold each Contributor harmless for any liability\n incurred by, or claims asserted against, such Contributor by reason\n of your accepting any such warranty or additional liability.\n\n END OF TERMS AND CONDITIONS\n\n APPENDIX: How to apply the Apache License to your work.\n\n To apply the Apache License to your work, attach the following\n boilerplate notice, with the fields enclosed by brackets \"[]\"\n replaced with your own identifying information. (Don't include\n the brackets!) The text should be enclosed in the appropriate\n comment syntax for the file format. We also recommend that a\n file or class name and description of purpose be included on the\n same \"printed page\" as the copyright notice for easier\n identification within third-party archives.\n\n Copyright [yyyy] [name of copyright owner]\n\n Licensed under the Apache License, Version 2.0 (the \"License\");\n you may not use this file except in compliance with the License.\n You may obtain a copy of the License at\n\n http://www.apache.org/licenses/LICENSE-2.0\n\n Unless required by applicable law or agreed to in writing, software\n distributed under the License is distributed on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n See the License for the specific language governing permissions and\n limitations under the License.\n"
},
{
"path": "MANIFEST.in",
"content": "include paddlespeech/t2s/exps/*.txt\ninclude paddlespeech/t2s/frontend/*.yaml"
},
{
"path": "README.md",
"content": "([简体中文](./README_cn.md)|English)\n\n ![](\"./docs/images/PaddleSpeech_logo.png\")
\n
\n\n\n ![](\"https://img.shields.io/badge/license-Apache%202-red.svg\")
\n ![](\"https://img.shields.io/github/v/release/PaddlePaddle/PaddleSpeech?color=ffa\")
\n ![](\"https://img.shields.io/badge/os-linux%2C%20win%2C%20mac-pink.svg\")
\n ![](\"https://img.shields.io/badge/python-3.8+-aff.svg\")
\n ![](\"https://img.shields.io/github/contributors/PaddlePaddle/PaddleSpeech?color=9ea\")
\n ![](\"https://img.shields.io/github/commit-activity/m/PaddlePaddle/PaddleSpeech?color=3af\")
\n ![](\"https://img.shields.io/github/issues/PaddlePaddle/PaddleSpeech?color=9cc\")
\n ![](\"https://img.shields.io/github/stars/PaddlePaddle/PaddleSpeech?color=ccf\")
\n ![](\"https://img.shields.io/pypi/dm/PaddleSpeech\")
\n ![](\"https://static.pepy.tech/badge/paddlespeech\")
\n ![](\"https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue\")
\n
\n\n\n------------------------------------------------------------------------------------\n\n**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for a variety of critical tasks in speech and audio, with the state-of-art and influential models. \n\n**PaddleSpeech** won the [NAACL2022 Best Demo Award](https://2022.naacl.org/blog/best-demo-award/), please check out our paper on [Arxiv](https://arxiv.org/abs/2205.12007).\n\n##### Speech Recognition\n\n\n
\n \n \n | Input Audio | \n Recognition Result | \n
\n \n \n \n \n \n ![](\"./docs/images/audio_icon.png\")
\n | \n I knocked at the door on the ancient side of the building. | \n
\n \n \n \n
\n | \n 我认为跑步最重要的就是给我带来了身体健康。 | \n
\n \n
\n\n
\n
\n \n \n | Input Audio | \n Translations Result | \n
\n \n \n \n \n \n
\n | \n 我 在 这栋 建筑 的 古老 门上 敲门。 | \n
\n \n
\n\n
\n
\n \n \n | Input Text | \n Output Text | \n
\n \n \n \n | 今天的天气真不错啊你下午有空吗我想约你一起去吃饭 | \n 今天的天气真不错啊！你下午有空吗？我想约你一起去吃饭。 | \n
\n \n
\n\n
\n
![](\"https://user-images.githubusercontent.com/30135920/212860467-9e943cc3-8be8-49a4-97fd-7c94aad8e979.jpg\")
\n
（Click to expand）Open Source Speech Recognition
\n\n**command line experience**\n\n```shell\npaddlespeech asr --lang zh --input zh.wav\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.asr.infer import ASRExecutor\n>>> asr = ASRExecutor()\n>>> result = asr(audio_file=\"zh.wav\")\n>>> print(result)\n我认为跑步最重要的就是给我带来了身体健康\n```\n Open Source Speech Synthesis
\n\nOutput 24k sample rate wav format audio\n\n\n**command line experience**\n\n```shell\npaddlespeech tts --input \"你好，欢迎使用百度飞桨深度学习框架！\" --output output.wav\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.tts.infer import TTSExecutor\n>>> tts = TTSExecutor()\n>>> tts(text=\"今天天气十分不错。\", output=\"output.wav\")\n```\n- You can experience in [Huggingface Spaces](https://huggingface.co/spaces) [TTS Demo](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS)\n\n An open-domain sound classification tool
\n\nSound classification model based on 527 categories of AudioSet dataset\n\n**command line experience**\n\n```shell\npaddlespeech cls --input zh.wav\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.cls.infer import CLSExecutor\n>>> cls = CLSExecutor()\n>>> result = cls(audio_file=\"zh.wav\")\n>>> print(result)\nSpeech 0.9027186632156372\n```\n\n Industrial-grade voiceprint extraction tool
\n\n**command line experience**\n\n```shell\npaddlespeech vector --task spk --input zh.wav\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.vector import VectorExecutor\n>>> vec = VectorExecutor()\n>>> result = vec(audio_file=\"zh.wav\")\n>>> print(result) # 187维向量\n[ -0.19083306 9.474295 -14.122263 -2.0916545 0.04848729\n 4.9295826 1.4780062 0.3733844 10.695862 3.2697146\n -4.48199 -0.6617882 -9.170393 -11.1568775 -1.2358263 ...]\n```\n\n Quick recovery of text punctuation, works with ASR models
\n\n**command line experience**\n\n```shell\npaddlespeech text --task punc --input 今天的天气真不错啊你下午有空吗我想约你一起去吃饭\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.text.infer import TextExecutor\n>>> text_punc = TextExecutor()\n>>> result = text_punc(text=\"今天的天气真不错啊你下午有空吗我想约你一起去吃饭\")\n今天的天气真不错啊！你下午有空吗？我想约你一起去吃饭。\n```\n\n End-to-end English to Chinese Speech Translation Tool
\n\nUse pre-compiled kaldi related tools, only support experience in Ubuntu system\n\n**command line experience**\n\n```shell\npaddlespeech st --input en.wav\n```\n\n**Python API experience**\n\n```python\n>>> from paddlespeech.cli.st.infer import STExecutor\n>>> st = STExecutor()\n>>> result = st(audio_file=\"en.wav\")\n['我 在 这栋 建筑 的 古老 门上 敲门 。']\n```\n\n\n \n \n | Task | \n Dataset | \n Model Type | \n Example | \n
\n \n \n \n | Audio Classification | \n ESC-50 | \n PANN | \n \n pann-esc50\n | \n
\n \n
\n\n\n\n**Keyword Spotting**\n\n\n \n \n | Task | \n Dataset | \n Model Type | \n Example | \n
\n \n \n \n | Keyword Spotting | \n hey-snips | \n MDTC | \n \n mdtc-hey-snips\n | \n
\n \n
\n\n\n\n**Speaker Verification**\n\n\n \n \n | Task | \n Dataset | \n Model Type | \n Example | \n
\n \n \n \n | Speaker Verification | \n VoxCeleb1/2 | \n ECAPA-TDNN | \n \n ecapa-tdnn-voxceleb12\n | \n
\n \n
\n\n\n\n**Speaker Diarization**\n\n\n \n \n | Task | \n Dataset | \n Model Type | \n Example | \n
\n \n \n \n | Speaker Diarization | \n AMI | \n ECAPA-TDNN + AHC / SC | \n \n ecapa-tdnn-ami\n | \n
\n \n
\n\n\n\n**Punctuation Restoration**\n\n\n \n \n | Task | \n Dataset | \n Model Type | \n Example | \n
\n \n \n \n | Punctuation Restoration | \n IWLST2012_zh | \n Ernie Linear | \n \n iwslt2012-punc0\n | \n
\n \n
\n\n## Documents\n\nNormally, [Speech SoTA](https://paperswithcode.com/area/speech), [Audio SoTA](https://paperswithcode.com/area/audio) and [Music SoTA](https://paperswithcode.com/area/music) give you an overview of the hot academic topics in the related area. To focus on the tasks in PaddleSpeech, you will find the following guidelines are helpful to grasp the core ideas.\n\n- [Installation](./docs/source/install.md)\n- [Quick Start](#quickstart)\n- [Some Demos](./demos/README.md)\n- Tutorials\n - [Automatic Speech Recognition](./docs/source/asr/quick_start.md)\n - [Introduction](./docs/source/asr/models_introduction.md)\n - [Data Preparation](./docs/source/asr/data_preparation.md)\n - [Ngram LM](./docs/source/asr/ngram_lm.md)\n - [Text-to-Speech](./docs/source/tts/quick_start.md)\n - [Introduction](./docs/source/tts/models_introduction.md)\n - [Advanced Usage](./docs/source/tts/advanced_usage.md)\n - [Chinese Rule Based Text Frontend](./docs/source/tts/zh_text_frontend.md)\n - [Test Audio Samples](https://paddlespeech.readthedocs.io/en/latest/tts/demo.html)\n - Speaker Verification\n - [Audio Searching](./demos/audio_searching/README.md)\n - [Speaker Verification](./demos/speaker_verification/README.md)\n - [Audio Classification](./demos/audio_tagging/README.md)\n - [Speech Translation](./demos/speech_translation/README.md)\n - [Speech Server](./demos/speech_server/README.md)\n- [Released Models](./docs/source/released_model.md)\n - [Speech-to-Text](#SpeechToText)\n - [Text-to-Speech](#TextToSpeech)\n - [Audio Classification](#AudioClassification)\n - [Speaker Verification](#SpeakerVerification)\n - [Speaker Diarization](#SpeakerDiarization)\n - [Punctuation Restoration](#PunctuationRestoration)\n- [Community](#Community)\n- [Welcome to contribute](#contribution)\n- [License](#License)\n\nThe Text-to-Speech module is originally called [Parakeet](https://github.com/PaddlePaddle/Parakeet), and now merged with this repository. If you are interested in academic research about this task, please see [TTS research overview](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/docs/source/tts#overview). Also, [this document](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/tts/models_introduction.md) is a good guideline for the pipeline components.\n\n\n## ⭐ Examples\n- **[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo): Use PaddleSpeech TTS to generate virtual human voice.**\n \n\n\n- [PaddleSpeech Demo Video](https://paddlespeech.readthedocs.io/en/latest/demo_video.html)\n\n- **[VTuberTalk](https://github.com/jerryuhoo/VTuberTalk): Use PaddleSpeech TTS and ASR to clone voice from videos.**\n\n\n## Citation\n\nTo cite PaddleSpeech for research, please use the following format.\n\n```text\n@inproceedings{zhang2022paddlespeech,\n title = {PaddleSpeech: An Easy-to-Use All-in-One Speech Toolkit},\n author = {Hui Zhang, Tian Yuan, Junkun Chen, Xintong Li, Renjie Zheng, Yuxin Huang, Xiaojie Chen, Enlei Gong, Zeyu Chen, Xiaoguang Hu, dianhai yu, Yanjun Ma, Liang Huang},\n booktitle = {Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Demonstrations},\n year = {2022},\n publisher = {Association for Computational Linguistics},\n}\n\n@InProceedings{pmlr-v162-bai22d,\n title = {{A}$^3${T}: Alignment-Aware Acoustic and Text Pretraining for Speech Synthesis and Editing},\n author = {Bai, He and Zheng, Renjie and Chen, Junkun and Ma, Mingbo and Li, Xintong and Huang, Liang},\n booktitle = {Proceedings of the 39th International Conference on Machine Learning},\n pages = {1399--1411},\n year = {2022},\n volume = {162},\n series = {Proceedings of Machine Learning Research},\n month = {17--23 Jul},\n publisher = {PMLR},\n pdf = {https://proceedings.mlr.press/v162/bai22d/bai22d.pdf},\n url = {https://proceedings.mlr.press/v162/bai22d.html},\n}\n\n@inproceedings{zheng2021fused,\n title={Fused acoustic and text encoding for multimodal bilingual pretraining and speech translation},\n author={Zheng, Renjie and Chen, Junkun and Ma, Mingbo and Huang, Liang},\n booktitle={International Conference on Machine Learning},\n pages={12736--12746},\n year={2021},\n organization={PMLR}\n}\n```\n\n\n## Contribute to PaddleSpeech\n\nYou are warmly welcome to submit questions in [discussions](https://github.com/PaddlePaddle/PaddleSpeech/discussions) and bug reports in [issues](https://github.com/PaddlePaddle/PaddleSpeech/issues)! Also, we highly appreciate if you are willing to contribute to this project!\n\n### Contributors\n\n![](\"https://avatars.githubusercontent.com/u/3038472?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/87408988?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/24568452?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/11361692?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/22954146?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/3064195?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/34430015?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/56786796?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/7038341?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/5782283?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/10339970?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/3139179?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/23690325?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/92848346?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/20623194?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/24245709?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/97937340?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/18379485?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/30135920?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/6771821?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/48039828?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/38800877?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/24390500?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/6836917?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/7534971?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/54951765?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/51190264?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/15189190?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/91456992?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/29157444?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/16222986?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/46859427?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/23568094?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/1371212?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/101700995?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/81195143?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/7845005?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/23097963?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/22723154?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/4719594?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/61459181?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/32832641?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/20540661?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/in/29110?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/6442159?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/34813030?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/97284124?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/96406262?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/25839309?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/82639260?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/523300?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/46156734?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/21351065?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/52739577?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/16233945?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/61349199?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/71760778?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/34057289?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/58069375?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/6285069?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/73375426?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/2288870?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/6081200?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/2833376?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/1307053?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/10704539?s=60&v=4\")
\n![](\"https://avatars.githubusercontent.com/u/18050469?s=60&v=4\")
\n
\n\n## Acknowledgement\n- Many thanks to [HighCWu](https://github.com/HighCWu) for adding [VITS-aishell3](./examples/aishell3/vits) and [VITS-VC](./examples/aishell3/vits-vc) examples.\n- Many thanks to [david-95](https://github.com/david-95) for fixing multi-punctuation bug、contributing to multiple program and data, and adding [SSML](https://github.com/PaddlePaddle/PaddleSpeech/discussions/2538) for TTS Chinese Text Frontend. \n- Many thanks to [BarryKCL](https://github.com/BarryKCL) for improving TTS Chinses Frontend based on [G2PW](https://github.com/GitYCC/g2pW).\n- Many thanks to [yeyupiaoling](https://github.com/yeyupiaoling)/[PPASR](https://github.com/yeyupiaoling/PPASR)/[PaddlePaddle-DeepSpeech](https://github.com/yeyupiaoling/PaddlePaddle-DeepSpeech)/[VoiceprintRecognition-PaddlePaddle](https://github.com/yeyupiaoling/VoiceprintRecognition-PaddlePaddle)/[AudioClassification-PaddlePaddle](https://github.com/yeyupiaoling/AudioClassification-PaddlePaddle) for years of attention, constructive advice and great help.\n- Many thanks to [mymagicpower](https://github.com/mymagicpower) for the Java implementation of ASR upon [short](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_sdk) and [long](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_long_audio_sdk) audio files.\n- Many thanks to [JiehangXie](https://github.com/JiehangXie)/[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo) for developing Virtual Uploader(VUP)/Virtual YouTuber(VTuber) with PaddleSpeech TTS function.\n- Many thanks to [745165806](https://github.com/745165806)/[PaddleSpeechTask](https://github.com/745165806/PaddleSpeechTask) for contributing Punctuation Restoration model.\n- Many thanks to [kslz](https://github.com/745165806) for supplementary Chinese documents.\n- Many thanks to [awmmmm](https://github.com/awmmmm) for contributing fastspeech2 aishell3 conformer pretrained model.\n- Many thanks to [phecda-xu](https://github.com/phecda-xu)/[PaddleDubbing](https://github.com/phecda-xu/PaddleDubbing) for developing a dubbing tool with GUI based on PaddleSpeech TTS model.\n- Many thanks to [jerryuhoo](https://github.com/jerryuhoo)/[VTuberTalk](https://github.com/jerryuhoo/VTuberTalk) for developing a GUI tool based on PaddleSpeech TTS and code for making datasets from videos based on PaddleSpeech ASR.\n- Many thanks to [vpegasus](https://github.com/vpegasus)/[xuesebot](https://github.com/vpegasus/xuesebot) for developing a rasa chatbot,which is able to speak and listen thanks to PaddleSpeech.\n- Many thanks to [chenkui164](https://github.com/chenkui164)/[FastASR](https://github.com/chenkui164/FastASR) for the C++ inference implementation of PaddleSpeech ASR.\n- Many thanks to [heyudage](https://github.com/heyudage)/[VoiceTyping](https://github.com/heyudage/VoiceTyping) for the real-time voice typing tool implementation of PaddleSpeech ASR streaming services.\n- Many thanks to [EscaticZheng](https://github.com/EscaticZheng)/[ps3.9wheel-install](https://github.com/EscaticZheng/ps3.9wheel-install) for the python3.9 prebuilt wheel for PaddleSpeech installation in Windows without Visual Studio.\nBesides, PaddleSpeech depends on a lot of open source repositories. See [references](./docs/source/reference.md) for more information.\n- Many thanks to [chinobing](https://github.com/chinobing)/[FastAPI-PaddleSpeech-Audio-To-Text](https://github.com/chinobing/FastAPI-PaddleSpeech-Audio-To-Text) for converting audio to text based on FastAPI and PaddleSpeech.\n- Many thanks to [MistEO](https://github.com/MistEO)/[Pallas-Bot](https://github.com/MistEO/Pallas-Bot) for QQ bot based on PaddleSpeech TTS.\n\n\n## License\n\nPaddleSpeech is provided under the [Apache-2.0 License](./LICENSE).\n\n## Stargazers over time\n\n[](https://starchart.cc/PaddlePaddle/PaddleSpeech)\n"
},
{
"path": "README_cn.md",
"content": "(简体中文|[English](./README.md))\n\n \n
\n\n\n\n \n \n \n \n \n \n \n \n \n \n \n
\n\n\n\n------------------------------------------------------------------------------------\n\n**PaddleSpeech** 是基于飞桨 [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) 的语音方向的开源模型库，用于语音和音频中的各种关键任务的开发，包含大量基于深度学习前沿和有影响力的模型，一些典型的应用示例如下：\n\n**PaddleSpeech** 荣获 [NAACL2022 Best Demo Award](https://2022.naacl.org/blog/best-demo-award/), 请访问 [Arxiv](https://arxiv.org/abs/2205.12007) 论文。\n \n### 效果展示\n\n##### 语音识别\n\n\n
\n \n \n | 输入音频 | \n 识别结果 | \n
\n \n \n \n \n \n
\n | \n I knocked at the door on the ancient side of the building. | \n
\n \n \n \n
\n | \n 我认为跑步最重要的就是给我带来了身体健康。 | \n
\n \n
\n\n
\n
\n \n \n | 输入音频 | \n 翻译结果 | \n
\n \n \n \n \n \n
\n | \n 我 在 这栋 建筑 的 古老 门上 敲门。 | \n
\n \n
\n\n
\n
\n \n \n | 输入文本 | \n 输出文本 | \n
\n \n \n \n | 今天的天气真不错啊你下午有空吗我想约你一起去吃饭 | \n 今天的天气真不错啊！你下午有空吗？我想约你一起去吃饭。 | \n
\n \n
\n\n
\n
\n
（点击可展开）开源中文语音识别
\n\n命令行一键体验\n\n```shell\npaddlespeech asr --lang zh --input zh.wav\n```\n\nPython API 一键预测\n\n```python\n>>> from paddlespeech.cli.asr.infer import ASRExecutor\n>>> asr = ASRExecutor()\n>>> result = asr(audio_file=\"zh.wav\")\n>>> print(result)\n我认为跑步最重要的就是给我带来了身体健康\n```\n 开源中文语音合成
\n\n输出 24k 采样率wav格式音频\n\n\n命令行一键体验\n\n```shell\npaddlespeech tts --input \"你好，欢迎使用百度飞桨深度学习框架！\" --output output.wav\n```\n\nPython API 一键预测\n\n```python\n>>> from paddlespeech.cli.tts.infer import TTSExecutor\n>>> tts = TTSExecutor()\n>>> tts(text=\"今天天气十分不错。\", output=\"output.wav\")\n```\n- 语音合成的 web demo 已经集成进了 [Huggingface Spaces](https://huggingface.co/spaces). 请参考: [TTS Demo](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS)\n\n 适配多场景的开放领域声音分类工具
\n\n基于 AudioSet 数据集 527 个类别的声音分类模型\n\n命令行一键体验\n\n```shell\npaddlespeech cls --input zh.wav\n```\n\npython API 一键预测\n\n```python\n>>> from paddlespeech.cli.cls.infer import CLSExecutor\n>>> cls = CLSExecutor()\n>>> result = cls(audio_file=\"zh.wav\")\n>>> print(result)\nSpeech 0.9027186632156372\n```\n\n 工业级声纹提取工具
\n\n命令行一键体验\n\n```shell\npaddlespeech vector --task spk --input zh.wav\n```\n\nPython API 一键预测\n\n```python\n>>> from paddlespeech.cli.vector import VectorExecutor\n>>> vec = VectorExecutor()\n>>> result = vec(audio_file=\"zh.wav\")\n>>> print(result) # 187维向量\n[ -0.19083306 9.474295 -14.122263 -2.0916545 0.04848729\n 4.9295826 1.4780062 0.3733844 10.695862 3.2697146\n -4.48199 -0.6617882 -9.170393 -11.1568775 -1.2358263 ...]\n```\n\n 一键恢复文本标点，可与ASR模型配合使用
\n\n命令行一键体验\n\n```shell\npaddlespeech text --task punc --input 今天的天气真不错啊你下午有空吗我想约你一起去吃饭\n```\n\nPython API 一键预测\n\n```python\n>>> from paddlespeech.cli.text.infer import TextExecutor\n>>> text_punc = TextExecutor()\n>>> result = text_punc(text=\"今天的天气真不错啊你下午有空吗我想约你一起去吃饭\")\n今天的天气真不错啊！你下午有空吗？我想约你一起去吃饭。\n```\n\n 端到端英译中语音翻译工具
\n\n使用预编译的 kaldi 相关工具，只支持在 Ubuntu 系统中体验\n\n命令行一键体验\n\n```shell\npaddlespeech st --input en.wav\n```\n\npython API 一键预测\n\n```python\n>>> from paddlespeech.cli.st.infer import STExecutor\n>>> st = STExecutor()\n>>> result = st(audio_file=\"en.wav\")\n['我 在 这栋 建筑 的 古老 门上 敲门 。']\n```\n\n\n \n \n | 任务 | \n 数据集 | \n 模型类型 | \n 脚本 | \n
\n \n \n \n | 声音分类 | \n ESC-50 | \n PANN | \n \n pann-esc50\n | \n
\n \n
\n\n\n\n\n**语音唤醒**\n\n\n \n \n | 任务 | \n 数据集 | \n 模型类型 | \n 脚本 | \n
\n \n \n \n | 语音唤醒 | \n hey-snips | \n MDTC | \n \n mdtc-hey-snips\n | \n
\n \n
\n\n\n\n**声纹识别**\n\n\n \n \n | 任务 | \n 数据集 | \n 模型类型 | \n 脚本 | \n
\n \n \n \n | 声纹识别 | \n VoxCeleb1/2 | \n ECAPA-TDNN | \n \n ecapa-tdnn-voxceleb12\n | \n
\n \n
\n\n\n\n**说话人日志**\n\n\n \n \n | 任务 | \n 数据集 | \n 模型类型 | \n 脚本 | \n
\n \n \n \n | 说话人日志 | \n AMI | \n ECAPA-TDNN + AHC / SC | \n \n ecapa-tdnn-ami\n | \n
\n \n
\n\n\n\n**标点恢复**\n\n\n \n \n | 任务 | \n 数据集 | \n 模型类型 | \n 脚本 | \n
\n \n \n \n | 标点恢复 | \n IWLST2012_zh | \n Ernie Linear | \n \n iwslt2012-punc0\n | \n
\n \n
\n\n\n## 教程文档\n\n对于 PaddleSpeech 的所关注的任务，以下指南有助于帮助开发者快速入门，了解语音相关核心思想。\n\n- [下载安装](./docs/source/install_cn.md)\n- [快速开始](#快速开始)\n- Notebook基础教程\n - [声音分类](./docs/tutorial/cls/cls_tutorial.ipynb)\n - [语音识别](./docs/tutorial/asr/tutorial_transformer.ipynb)\n - [语音翻译](./docs/tutorial/st/st_tutorial.ipynb)\n - [声音合成](./docs/tutorial/tts/tts_tutorial.ipynb)\n - [示例Demo](./demos/README.md)\n- 进阶文档 \n - [语音识别自定义训练](./docs/source/asr/quick_start.md)\n - [简介](./docs/source/asr/models_introduction.md)\n - [数据准备](./docs/source/asr/data_preparation.md)\n - [Ngram 语言模型](./docs/source/asr/ngram_lm.md)\n - [语音合成自定义训练](./docs/source/tts/quick_start.md)\n - [简介](./docs/source/tts/models_introduction.md)\n - [进阶用法](./docs/source/tts/advanced_usage.md)\n - [中文文本前端](./docs/source/tts/zh_text_frontend.md)\n - [测试语音样本](https://paddlespeech.readthedocs.io/en/latest/tts/demo.html)\n - 声纹识别\n - [声纹识别](./demos/speaker_verification/README_cn.md)\n - [音频检索](./demos/audio_searching/README_cn.md)\n - [声音分类](./demos/audio_tagging/README_cn.md)\n - [语音翻译](./demos/speech_translation/README_cn.md)\n - [服务化部署](./demos/speech_server/README_cn.md)\n- [模型列表](#模型列表)\n - [语音识别](#语音识别模型)\n - [语音合成](#语音合成模型)\n - [声音分类](#声音分类模型)\n - [声纹识别](#声纹识别模型)\n - [说话人日志](#说话人日志模型)\n - [标点恢复](#标点恢复模型)\n- [技术交流群](#技术交流群)\n- [欢迎贡献](#欢迎贡献)\n- [License](#License)\n\n\n语音合成模块最初被称为 [Parakeet](https://github.com/PaddlePaddle/Parakeet)，现在与此仓库合并。如果您对该任务的学术研究感兴趣，请参阅 [TTS 研究概述](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/docs/source/tts#overview)。此外，[模型介绍](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/tts/models_introduction.md) 是了解语音合成流程的一个很好的指南。\n\n\n## ⭐ 应用案例\n- **[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo): 使用 PaddleSpeech 的语音合成模块生成虚拟人的声音。**\n \n\n\n- [PaddleSpeech 示例视频](https://paddlespeech.readthedocs.io/en/latest/demo_video.html)\n\n\n- **[VTuberTalk](https://github.com/jerryuhoo/VTuberTalk): 使用 PaddleSpeech 的语音合成和语音识别从视频中克隆人声。**\n\n\n
![](\"https://raw.githubusercontent.com/jerryuhoo/VTuberTalk/main/gui/gui.png\")
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\n## 致谢\n- 非常感谢 [HighCWu](https://github.com/HighCWu) 新增 [VITS-aishell3](./examples/aishell3/vits) 和 [VITS-VC](./examples/aishell3/vits-vc) 代码示例。\n- 非常感谢 [david-95](https://github.com/david-95) 修复 TTS 句尾多标点符号出错的问题，贡献补充多条程序和数据。为 TTS 中文文本前端新增 [SSML](https://github.com/PaddlePaddle/PaddleSpeech/discussions/2538) 功能。\n- 非常感谢 [BarryKCL](https://github.com/BarryKCL) 基于 [G2PW](https://github.com/GitYCC/g2pW) 对 TTS 中文文本前端的优化。\n- 非常感谢 [yeyupiaoling](https://github.com/yeyupiaoling)/[PPASR](https://github.com/yeyupiaoling/PPASR)/[PaddlePaddle-DeepSpeech](https://github.com/yeyupiaoling/PaddlePaddle-DeepSpeech)/[VoiceprintRecognition-PaddlePaddle](https://github.com/yeyupiaoling/VoiceprintRecognition-PaddlePaddle)/[AudioClassification-PaddlePaddle](https://github.com/yeyupiaoling/AudioClassification-PaddlePaddle) 多年来的关注和建议，以及在诸多问题上的帮助。\n- 非常感谢 [mymagicpower](https://github.com/mymagicpower) 采用PaddleSpeech 对 ASR 的[短语音](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_sdk)及[长语音](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_long_audio_sdk)进行 Java 实现。\n- 非常感谢 [JiehangXie](https://github.com/JiehangXie)/[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo) 采用 PaddleSpeech 语音合成功能实现 Virtual Uploader(VUP)/Virtual YouTuber(VTuber) 虚拟主播。\n- 非常感谢 [745165806](https://github.com/745165806)/[PaddleSpeechTask](https://github.com/745165806/PaddleSpeechTask) 贡献标点重建相关模型。\n- 非常感谢 [kslz](https://github.com/kslz) 补充中文文档。\n- 非常感谢 [awmmmm](https://github.com/awmmmm) 提供 fastspeech2 aishell3 conformer 预训练模型。\n- 非常感谢 [phecda-xu](https://github.com/phecda-xu)/[PaddleDubbing](https://github.com/phecda-xu/PaddleDubbing) 基于 PaddleSpeech 的 TTS 模型搭建带 GUI 操作界面的配音工具。\n- 非常感谢 [jerryuhoo](https://github.com/jerryuhoo)/[VTuberTalk](https://github.com/jerryuhoo/VTuberTalk) 基于 PaddleSpeech 的 TTS GUI 界面和基于 ASR 制作数据集的相关代码。\n- 非常感谢 [vpegasus](https://github.com/vpegasus)/[xuesebot](https://github.com/vpegasus/xuesebot) 基于 PaddleSpeech 的 ASR 与 TTS 设计的可听、说对话机器人。\n- 非常感谢 [chenkui164](https://github.com/chenkui164)/[FastASR](https://github.com/chenkui164/FastASR) 对 PaddleSpeech 的 ASR 进行 C++ 推理实现。\n- 非常感谢 [heyudage](https://github.com/heyudage)/[VoiceTyping](https://github.com/heyudage/VoiceTyping) 基于 PaddleSpeech 的 ASR 流式服务实现的实时语音输入法工具。\n- 非常感谢 [EscaticZheng](https://github.com/EscaticZheng)/[ps3.9wheel-install](https://github.com/EscaticZheng/ps3.9wheel-install) 对PaddleSpeech在Windows下的安装提供了无需Visua Studio，基于python3.9的预编译依赖安装包。\n- 非常感谢 [chinobing](https://github.com/chinobing)/[FastAPI-PaddleSpeech-Audio-To-Text](https://github.com/chinobing/FastAPI-PaddleSpeech-Audio-To-Text) 利用 FastAPI 实现 PaddleSpeech 语音转文字，文件上传、分割、转换进度显示、后台更新任务并以 csv 格式输出。\n- 非常感谢 [MistEO](https://github.com/MistEO)/[Pallas-Bot](https://github.com/MistEO/Pallas-Bot) 基于 PaddleSpeech TTS 的 QQ Bot 项目。\n\n此外，PaddleSpeech 依赖于许多开源存储库。有关更多信息，请参阅 [references](./docs/source/reference.md)。\n\n## License\n\nPaddleSpeech 在 [Apache-2.0 许可](./LICENSE) 下提供。\n\n## Stargazers over time\n\n[](https://starchart.cc/PaddlePaddle/PaddleSpeech)\n"
},
{
"path": "audio/CMakeLists.txt",
"content": "cmake_minimum_required(VERSION 3.16 FATAL_ERROR)\n\n# Use compiler ID \"AppleClang\" instead of \"Clang\" for XCode.\n# Not setting this sometimes makes XCode C compiler gets detected as \"Clang\",\n# even when the C++ one is detected as \"AppleClang\".\ncmake_policy(SET CMP0010 NEW)\ncmake_policy(SET CMP0025 NEW)\n\n# Suppress warning flags in default MSVC configuration. It's not\n# mandatory that we do this (and we don't if cmake is old), but it's\n# nice when it's possible, and it's possible on our Windows configs.\nif(NOT CMAKE_VERSION VERSION_LESS 3.15.0)\n cmake_policy(SET CMP0092 NEW)\nendif()\n\nproject(paddleaudio)\n\n# check and set CMAKE_CXX_STANDARD\nstring(FIND \"${CMAKE_CXX_FLAGS}\" \"-std=c++\" env_cxx_standard)\nif(env_cxx_standard GREATER -1)\n message(\n WARNING \"C++ standard version definition detected in environment variable.\"\n \"paddleaudio requires -std=c++14. Please remove -std=c++ settings in your environment.\")\nendif()\n\n\nset(CMAKE_CXX_STANDARD 14)\nset(CMAKE_C_STANDARD 11)\n\nset(CMAKE_EXPORT_COMPILE_COMMANDS ON)\nset(CMAKE_POSITION_INDEPENDENT_CODE ON)\nset(CMAKE_VERBOSE_MAKEFILE ON)\n\n# Options\noption(BUILD_SOX \"Build libsox statically\" ON)\noption(BUILD_MAD \"Enable libmad\" ON)\noption(BUILD_KALDI \"Build kaldi statically\" ON)\noption(BUILD_PADDLEAUDIO_PYTHON_EXTENSION \"Build Python extension\" ON)\n\n\n# cmake\nset(CMAKE_MODULE_PATH \"${CMAKE_MODULE_PATH};${PROJECT_SOURCE_DIR}/cmake;${PROJECT_SOURCE_DIR}/cmake/external\")\n\n# fc_patch dir\nset(FETCHCONTENT_QUIET off)\nget_filename_component(fc_patch \"fc_patch\" REALPATH BASE_DIR \"${CMAKE_SOURCE_DIR}\")\nset(FETCHCONTENT_BASE_DIR ${fc_patch})\nset(THIRD_PARTY_PATH ${fc_patch})\n\nset(PYBIND11_PYTHON_VERSION ${PY_VERSION})\ninclude(cmake/pybind.cmake)\ninclude_directories(${PYTHON_INCLUDE_DIR})\n\ninclude_directories(${CMAKE_CURRENT_SOURCE_DIR}/paddleaudio/third_party/)\n\n# packages\nfind_package(Python3 COMPONENTS Interpreter Development)\n\n# set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -pthread -O0 -Wall -g\")\nadd_subdirectory(paddleaudio)\n\n# Summary\ninclude(cmake/summary.cmake)\nonnx_print_configuration_summary()\n"
},
{
"path": "audio/README.md",
"content": "# PaddleAudio\n\n安装方式： pip install paddleaudio\n\n目前支持的平台：Linux, Mac, Windows\n\n## Environment\n\n## Build wheel\ncmd: python setup.py bdist_wheel\n\nLinux test build whl environment:\n* os - Ubuntu 16.04.7 LTS\n* gcc/g++ - 8.2.0\n* cmake - 3.18.0 (need install)\n\nMAC：test build whl environment：\n* os \n* gcc/g++ 12.2.0\n* cpu Intel Xeon E5 x86_64\n\nWindows：\nnot support paddleaudio C++ extension lib (sox io, kaldi native fbank)\n"
},
{
"path": "audio/cmake/FindGFortranLibs.cmake",
"content": "#.rst:\n# FindGFortranLibs\n# --------\n# https://github.com/Argonne-National-Laboratory/PIPS/blob/master/cmake/Modules/FindGFortranLibs.cmake\n# https://enccs.github.io/cmake-workshop/cxx-fortran/\n#\n# Find gcc Fortran compiler & library paths\n#\n# The module defines the following variables:\n#\n# ::\n#\n#\n# GFORTRANLIBS_FOUND - true if system has gfortran\n# LIBGFORTRAN_LIBRARIES - path to libgfortran\n# LIBQUADMATH_LIBRARIES - path to libquadmath\n# GFORTRAN_LIBARIES_DIR - directory containing libgfortran, libquadmath\n# GFORTRAN_INCLUDE_DIR - directory containing gfortran/gcc headers\n# LIBGOMP_LIBRARIES - path to libgomp\n# LIBGOMP_INCLUDE_DIR - directory containing omp.h header\n# GFORTRAN_VERSION_STRING - version of gfortran found\n#\nset(CMAKE_REQUIRED_QUIET ${LIBIOMP_FIND_QUIETLY})\n\nif(NOT CMAKE_REQUIRED_QUIET)\n message(STATUS \"Looking for gfortran related libraries...\")\nendif()\n\nenable_language(Fortran)\nif(CMAKE_Fortran_COMPILER_ID MATCHES \"GNU\")\n\n # Basically, call \"gfortran -v\" to dump compiler info to the string\n # GFORTRAN_VERBOSE_STR, which will be used to get necessary paths\n message(STATUS \"Extracting library and header information by calling 'gfortran -v'...\")\n execute_process(COMMAND \"${CMAKE_Fortran_COMPILER}\" \"-v\" ERROR_VARIABLE\n GFORTRAN_VERBOSE_STR RESULT_VARIABLE FLAG)\n\n # For debugging\n message(STATUS \"'gfortran -v' returned:\")\n message(STATUS \"${GFORTRAN_VERBOSE_STR}\")\n\n # Detect gfortran version\n string(REGEX MATCH \"gcc version [^\\t\\n ]+\" GFORTRAN_VER_STR \"${GFORTRAN_VERBOSE_STR}\")\n string(REGEX REPLACE \"gcc version ([^\\t\\n ]+)\" \"\\\\1\" GFORTRAN_VERSION_STRING \"${GFORTRAN_VER_STR}\")\n message(STATUS \"Detected gfortran version ${GFORTRAN_VERSION_STRING}\")\n unset(GFORTRAN_VER_STR)\n\n set(MATCH_REGEX \"[^\\t\\n ]+[\\t\\n ]+\")\n set(REPLACE_REGEX \"([^\\t\\n ]+)\")\n\n # Find architecture for compiler\n string(REGEX MATCH \"Target: [^\\t\\n ]+\"\n GFORTRAN_ARCH_STR \"${GFORTRAN_VERBOSE_STR}\")\n message(STATUS \"Architecture string: ${GFORTRAN_ARCH_STR}\")\n string(REGEX REPLACE \"Target: ([^\\t\\n ]+)\" \"\\\\1\"\n GFORTRAN_ARCH \"${GFORTRAN_ARCH_STR}\")\n message(STATUS \"Detected gfortran architecture: ${GFORTRAN_ARCH}\")\n unset(GFORTRAN_ARCH_STR)\n\n # Find install prefix, if it exists; if not, use default\n string(REGEX MATCH \"--prefix=[^\\t\\n ]+[\\t\\n ]+\"\n GFORTRAN_PREFIX_STR \"${GFORTRAN_VERBOSE_STR}\")\n if(NOT GFORTRAN_PREFIX_STR)\n message(STATUS \"Detected default gfortran prefix\")\n set(GFORTRAN_PREFIX_DIR \"/usr/local\") # default prefix for gcc install\n else()\n string(REGEX REPLACE \"--prefix=([^\\t\\n ]+)\" \"\\\\1\"\n GFORTRAN_PREFIX_DIR \"${GFORTRAN_PREFIX_STR}\")\n endif()\n message(STATUS \"Detected gfortran prefix: ${GFORTRAN_PREFIX_DIR}\")\n unset(GFORTRAN_PREFIX_STR)\n\n # Find install exec-prefix, if it exists; if not, use default\n string(REGEX MATCH \"--exec-prefix=[^\\t\\n ]+[\\t\\n ]+\" \"\\\\1\"\n GFORTRAN_EXEC_PREFIX_STR \"${GFORTRAN_VERBOSE_STR}\")\n if(NOT GFORTRAN_EXEC_PREFIX_STR)\n message(STATUS \"Detected default gfortran exec-prefix\")\n set(GFORTRAN_EXEC_PREFIX_DIR \"${GFORTRAN_PREFIX_DIR}\")\n else()\n string(REGEX REPLACE \"--exec-prefix=([^\\t\\n ]+)\" \"\\\\1\"\n GFORTRAN_EXEC_PREFIX_DIR \"${GFORTRAN_EXEC_PREFIX_STR}\")\n endif()\n message(STATUS \"Detected gfortran exec-prefix: ${GFORTRAN_EXEC_PREFIX_DIR}\")\n UNSET(GFORTRAN_EXEC_PREFIX_STR)\n\n # Find library directory and include directory, if library directory specified\n string(REGEX MATCH \"--libdir=[^\\t\\n ]+\"\n GFORTRAN_LIB_DIR_STR \"${GFORTRAN_VERBOSE_STR}\")\n if(NOT GFORTRAN_LIB_DIR_STR)\n message(STATUS \"Found --libdir flag -- not found\")\n message(STATUS \"Using default gfortran library & include directory paths\")\n string(STRIP ${GFORTRAN_PREFIX_DIR} TMPLIBDIR)\n set(GFORTRAN_LIBRARIES_DIR \"${TMPLIBDIR}/lib64\")\n set(GFORTRAN_INCLUDE_DIR \"${TMPLIBDIR}/include\")\n else()\n message(STATUS \"Found --libdir flag -- yes\")\n string(REGEX REPLACE \"--libdir=([^\\t\\n ]+)\" \"\\\\1\"\n GFORTRAN_LIBRARIES_DIR \"${GFORTRAN_LIB_DIR_STR}\")\n string(CONCAT GFORTRAN_INCLUDE_DIR \"${GFORTRAN_LIBRARIES_DIR}\" \"/gcc/\" \"${GFORTRAN_ARCH}\" \"/\" \"${GFORTRAN_VERSION_STRING}\" \"/include\")\n endif()\n message(STATUS \"gfortran libraries path: ${GFORTRAN_LIBRARIES_DIR}\")\n message(STATUS \"gfortran include path dir: ${GFORTRAN_INCLUDE_DIR}\")\n unset(GFORTRAN_LIB_DIR_STR)\n\n # There are lots of other build options for gcc & gfortran. For now, the\n # options implemented above should cover a lot of common use cases.\n\n # Clean up be deleting the output string from \"gfortran -v\"\n unset(GFORTRAN_VERBOSE_STR)\n\n # Find paths for libgfortran, libquadmath, libgomp\n # libgomp needed for OpenMP support without Clang\n find_library(LIBGFORTRAN_LIBRARIES NAMES gfortran libgfortran\n HINTS ${GFORTRAN_LIBRARIES_DIR})\n find_library(LIBQUADMATH_LIBRARIES NAMES quadmath libquadmath\n HINTS ${GFORTRAN_LIBRARIES_DIR})\n find_library(LIBGOMP_LIBRARIES NAMES gomp libgomp\n HINTS ${GFORTRAN_LIBRARIES_DIR})\n\n # Find OpenMP headers\n find_path(LIBGOMP_INCLUDE_DIR NAMES omp.h HINTS ${GFORTRAN_INCLUDE_DIR})\n\nelse()\n message(STATUS \"CMAKE_Fortran_COMPILER_ID does not match 'GNU'!\")\nendif()\n\ninclude(FindPackageHandleStandardArgs)\n\n# Required: libgfortran, libquadmath, path for gfortran libraries\n# Optional: libgomp, path for OpenMP headers, path for gcc/gfortran headers\nfind_package_handle_standard_args(GFortranLibs\n REQUIRED_VARS LIBGFORTRAN_LIBRARIES LIBQUADMATH_LIBRARIES GFORTRAN_LIBRARIES_DIR\n VERSION_VAR GFORTRAN_VERSION_STRING)\n\nif(GFORTRANLIBS_FOUND)\n message(STATUS \"Looking for gfortran libraries -- found\")\n message(STATUS \"gfortran version: ${GFORTRAN_VERSION_STRING}\")\nelse()\n message(STATUS \"Looking for gfortran libraries -- not found\")\nendif()\n\nmark_as_advanced(LIBGFORTRAN_LIBRARIES LIBQUADMATH_LIBRARIES\n LIBGOMP_LIBRARIES LIBGOMP_INCLUDE_DIR\n GFORTRAN_LIBRARIES_DIR GFORTRAN_INCLUDE_DIR)\n# FindGFortranLIBS.cmake ends here\n\n\nmessage(STATUS LIBGFORTRAN_LIBRARIES= ${LIBGFORTRAN_LIBRARIES})\nmessage(STATUS LIBQUADMATH_LIBRARIES= ${LIBQUADMATH_LIBRARIES})\nmessage(STATUS LIBGOMP_LIBRARIES= ${LIBGOMP_LIBRARIES})\nmessage(STATUS LIBGOMP_INCLUDE_DIR= ${LIBGOMP_INCLUDE_DIR})\nmessage(STATUS GFORTRAN_LIBRARIES_DIR= ${GFORTRAN_LIBRARIES_DIR})\nmessage(STATUS GFORTRAN_INCLUDE_DIR= ${GFORTRAN_INCLUDE_DIR})\n"
},
{
"path": "audio/cmake/external/openblas.cmake",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\ninclude(ExternalProject)\n\nset(CBLAS_PREFIX_DIR ${THIRD_PARTY_PATH}/openblas)\nset(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)\nset(CBLAS_REPOSITORY https://github.com/xianyi/OpenBLAS.git)\nset(CBLAS_TAG v0.3.10)\n\nif(NOT WIN32)\n set(CBLAS_LIBRARIES\n \"${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}\"\n CACHE FILEPATH \"openblas library.\" FORCE)\n set(CBLAS_INC_DIR\n \"${CBLAS_INSTALL_DIR}/include\"\n CACHE PATH \"openblas include directory.\" FORCE)\n set(OPENBLAS_CC\n \"${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable\")\n\n if(APPLE)\n set(OPENBLAS_CC \"${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}\")\n endif()\n set(OPTIONAL_ARGS \"\")\n set(COMMON_ARGS \"\")\n\n if(APPLE)\n if(CMAKE_SYSTEM_PROCESSOR MATCHES \"^x86(_64)?$\")\n set(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)\n endif()\n set(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1)\n endif()\n\n ExternalProject_Add(\n OPENBLAS\n URL \"https://paddleaudio.bj.bcebos.com/build/OpenBLAS-0.3.10.zip\"\n GIT_SHALLOW YES\n DOWNLOAD_DIR ${CBLAS_PREFIX_DIR}\n SOURCE_DIR ${CBLAS_PREFIX_DIR}\n INSTALL_DIR ${CBLAS_INSTALL_DIR}\n BUILD_IN_SOURCE 1\n BUILD_COMMAND make -j${NPROC} ${COMMON_ARGS} ${OPTIONAL_ARGS}\n INSTALL_COMMAND make install PREFIX=\n UPDATE_COMMAND \"\"\n CONFIGURE_COMMAND \"\"\n BUILD_BYPRODUCTS ${CBLAS_LIBRARIES})\n\n ExternalProject_Get_Property(OPENBLAS INSTALL_DIR)\n set(OpenBLAS_INSTALL_PREFIX ${INSTALL_DIR})\n add_library(openblas STATIC IMPORTED)\n add_dependencies(openblas OPENBLAS)\n set_target_properties(openblas PROPERTIES IMPORTED_LINK_INTERFACE_LANGUAGES Fortran)\n set_target_properties(openblas PROPERTIES IMPORTED_LOCATION ${OpenBLAS_INSTALL_PREFIX}/lib/libopenblas.a)\n\n link_directories(${OpenBLAS_INSTALL_PREFIX}/lib)\n include_directories(${OpenBLAS_INSTALL_PREFIX}/include)\n\n set(OPENBLAS_LIBRARIES\n ${OpenBLAS_INSTALL_PREFIX}/lib/libopenblas.a\n )\n\n add_library(libopenblas INTERFACE)\n add_dependencies(libopenblas openblas)\n target_include_directories(libopenblas INTERFACE ${OpenBLAS_INSTALL_PREFIX}/include/openblas)\n target_link_libraries(libopenblas INTERFACE ${OPENBLAS_LIBRARIES})\nelse()\n set(CBLAS_LIBRARIES\n \"${CBLAS_INSTALL_DIR}/lib/openblas${CMAKE_STATIC_LIBRARY_SUFFIX}\"\n CACHE FILEPATH \"openblas library.\" FORCE)\n set(CBLAS_INC_DIR\n \"${CBLAS_INSTALL_DIR}/include/openblas\"\n CACHE PATH \"openblas include directory.\" FORCE)\n ExternalProject_Add(\n extern_openblas\n ${EXTERNAL_PROJECT_LOG_ARGS}\n GIT_REPOSITORY ${CBLAS_REPOSITORY}\n GIT_TAG ${CBLAS_TAG}\n PREFIX ${CBLAS_PREFIX_DIR}\n INSTALL_DIR ${CBLAS_INSTALL_DIR}\n BUILD_IN_SOURCE 0\n UPDATE_COMMAND \"\"\n CMAKE_ARGS -DCMAKE_C_COMPILER=clang-cl\n -DCMAKE_CXX_COMPILER=clang-cl\n -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}\n -DCMAKE_INSTALL_PREFIX=${CBLAS_INSTALL_DIR}\n -DCMAKE_BUILD_TYPE=Release #${THIRD_PARTY_BUILD_TYPE}\n -DCMAKE_MT=mt\n -DUSE_THREAD=OFF\n -DBUILD_WITHOUT_LAPACK=NO\n -DCMAKE_Fortran_COMPILER=flang\n -DNOFORTRAN=0\n -DDYNAMIC_ARCH=ON\n #${EXTERNAL_OPTIONAL_ARGS}\n CMAKE_CACHE_ARGS\n -DCMAKE_INSTALL_PREFIX:PATH=${CBLAS_INSTALL_DIR}\n -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON\n -DCMAKE_BUILD_TYPE:STRING=Release #${THIRD_PARTY_BUILD_TYPE}\n # ninja need to know where openblas.lib comes from\n BUILD_BYPRODUCTS ${CBLAS_LIBRARIES})\n set(OPENBLAS_SHARED_LIB\n ${CBLAS_INSTALL_DIR}/bin/openblas${CMAKE_SHARED_LIBRARY_SUFFIX})\n\n add_library(openblas INTERFACE)\n add_dependencies(openblas extern_openblas)\n include_directories(${CBLAS_INC_DIR})\n link_libraries(${CBLAS_LIBRARIES})\nendif()\n\n"
},
{
"path": "audio/cmake/pybind.cmake",
"content": "#the pybind11 is from:https://github.com/pybind/pybind11\n# Copyright (c) 2016 Wenzel Jakob , All rights reserved.\n\nSET(PYBIND_ZIP \"v2.10.0.zip\")\nSET(LOCAL_PYBIND_ZIP ${FETCHCONTENT_BASE_DIR}/${PYBIND_ZIP})\nSET(PYBIND_SRC ${FETCHCONTENT_BASE_DIR}/pybind11)\nSET(DOWNLOAD_URL \"https://paddleaudio.bj.bcebos.com/build/v2.10.0.zip\")\nSET(PYBIND_TIMEOUT 600 CACHE STRING \"Timeout in seconds when downloading pybind.\")\n\nIF(NOT EXISTS ${LOCAL_PYBIND_ZIP})\n FILE(DOWNLOAD ${DOWNLOAD_URL}\n ${LOCAL_PYBIND_ZIP}\n TIMEOUT ${PYBIND_TIMEOUT}\n STATUS ERR\n SHOW_PROGRESS\n )\n\n IF(ERR EQUAL 0)\n MESSAGE(STATUS \"download pybind success\")\n ELSE()\n MESSAGE(FATAL_ERROR \"download pybind fail\")\n ENDIF()\nENDIF()\n\nIF(NOT EXISTS ${PYBIND_SRC})\n EXECUTE_PROCESS(\n COMMAND ${CMAKE_COMMAND} -E tar xfz ${LOCAL_PYBIND_ZIP}\n WORKING_DIRECTORY ${FETCHCONTENT_BASE_DIR}\n RESULT_VARIABLE tar_result\n )\n\n file(RENAME ${FETCHCONTENT_BASE_DIR}/pybind11-2.10.0 ${PYBIND_SRC})\n\n IF (tar_result MATCHES 0)\n MESSAGE(STATUS \"unzip pybind success\")\n ELSE()\n MESSAGE(FATAL_ERROR \"unzip pybind fail\")\n ENDIF()\n\nENDIF()\n\ninclude_directories(${PYBIND_SRC}/include)\n"
},
{
"path": "audio/cmake/summary.cmake",
"content": "# SPDX-License-Identifier: Apache-2.0\n\n# Prints accumulated ONNX configuration summary\nfunction (onnx_print_configuration_summary)\n message(STATUS \"\")\n message(STATUS \"******** Summary ********\")\n message(STATUS \" CMake version : ${CMAKE_VERSION}\")\n message(STATUS \" CMake command : ${CMAKE_COMMAND}\")\n message(STATUS \" System : ${CMAKE_SYSTEM_NAME}\")\n message(STATUS \" C++ compiler : ${CMAKE_CXX_COMPILER}\")\n message(STATUS \" C++ compiler version : ${CMAKE_CXX_COMPILER_VERSION}\")\n message(STATUS \" CXX flags : ${CMAKE_CXX_FLAGS}\")\n message(STATUS \" Build type : ${CMAKE_BUILD_TYPE}\")\n get_directory_property(tmp DIRECTORY ${PROJECT_SOURCE_DIR} COMPILE_DEFINITIONS)\n message(STATUS \" Compile definitions : ${tmp}\")\n message(STATUS \" CMAKE_PREFIX_PATH : ${CMAKE_PREFIX_PATH}\")\n message(STATUS \" CMAKE_INSTALL_PREFIX : ${CMAKE_INSTALL_PREFIX}\")\n message(STATUS \" CMAKE_MODULE_PATH : ${CMAKE_MODULE_PATH}\")\n message(STATUS \"\")\n message(STATUS \" ONNX version : ${ONNX_VERSION}\")\n message(STATUS \" ONNX NAMESPACE : ${ONNX_NAMESPACE}\")\n message(STATUS \" ONNX_USE_LITE_PROTO : ${ONNX_USE_LITE_PROTO}\")\n message(STATUS \" USE_PROTOBUF_SHARED_LIBS : ${ONNX_USE_PROTOBUF_SHARED_LIBS}\")\n message(STATUS \" Protobuf_USE_STATIC_LIBS : ${Protobuf_USE_STATIC_LIBS}\")\n message(STATUS \" ONNX_DISABLE_EXCEPTIONS : ${ONNX_DISABLE_EXCEPTIONS}\")\n message(STATUS \" ONNX_WERROR : ${ONNX_WERROR}\")\n message(STATUS \" ONNX_BUILD_TESTS : ${ONNX_BUILD_TESTS}\")\n message(STATUS \" ONNX_BUILD_BENCHMARKS : ${ONNX_BUILD_BENCHMARKS}\")\n message(STATUS \" ONNXIFI_DUMMY_BACKEND : ${ONNXIFI_DUMMY_BACKEND}\")\n message(STATUS \" ONNXIFI_ENABLE_EXT : ${ONNXIFI_ENABLE_EXT}\")\n message(STATUS \"\")\n message(STATUS \" Protobuf compiler : ${PROTOBUF_PROTOC_EXECUTABLE}\")\n message(STATUS \" Protobuf includes : ${PROTOBUF_INCLUDE_DIRS}\")\n message(STATUS \" Protobuf libraries : ${PROTOBUF_LIBRARIES}\")\n message(STATUS \" BUILD_ONNX_PYTHON : ${BUILD_ONNX_PYTHON}\")\n message(STATUS \" Python version : ${Python_VERSION}\")\n message(STATUS \" Python executable : ${Python_EXECUTABLE}\")\n message(STATUS \" Python includes : ${Python_INCLUDE_DIR}\")\n message(STATUS \" Python libraries : ${Python_LIBRARY}\")\n message(STATUS \" PYBIND11 : ${pybind11_FOUND}\")\n message(STATUS \" Pybind11 version : ${pybind11_VERSION}\")\n message(STATUS \" Pybind11 include : ${pybind11_INCLUDE_DIR}\")\n message(STATUS \" Pybind11 includes : ${pybind11_INCLUDE_DIRS}\")\n message(STATUS \" Pybind11 libraries : ${pybind11_LIBRARIES}\")\nendfunction()"
},
{
"path": "audio/paddleaudio/CMakeLists.txt",
"content": "\nadd_subdirectory(third_party)\nadd_subdirectory(src)\n"
},
{
"path": "audio/paddleaudio/__init__.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom . import _extension\nfrom . import backends\nfrom . import compliance\nfrom . import datasets\nfrom . import features\nfrom . import functional\nfrom . import metric\nfrom . import sox_effects\nfrom . import utils\n"
},
{
"path": "audio/paddleaudio/_extension.py",
"content": "import contextlib\nimport ctypes\nimport os\nimport sys\nimport types\nimport warnings\nfrom pathlib import Path\n\nfrom ._internal import module_utils as _mod_utils # noqa: F401\n\n# Query `hasattr` only once.\n_SET_GLOBAL_FLAGS = hasattr(sys, 'getdlopenflags') and hasattr(sys,\n 'setdlopenflags')\n\n\n@contextlib.contextmanager\ndef dl_open_guard():\n \"\"\"\n # https://manpages.debian.org/bullseye/manpages-dev/dlopen.3.en.html\n Context manager to set the RTLD_GLOBAL dynamic linker flag while we open a\n shared library to load custom operators.\n \"\"\"\n if _SET_GLOBAL_FLAGS:\n old_flags = sys.getdlopenflags()\n sys.setdlopenflags(old_flags | ctypes.RTLD_GLOBAL)\n yield\n if _SET_GLOBAL_FLAGS:\n sys.setdlopenflags(old_flags)\n\n\ndef resolve_library_path(path: str) -> str:\n return os.path.realpath(path)\n\n\nclass _Ops(types.ModuleType):\n #__file__ = '_ops.py'\n\n def __init__(self):\n super(_Ops, self).__init__('paddleaudio.ops')\n self.loaded_libraries = set()\n\n def load_library(self, path):\n \"\"\"\n Loads a shared library from the given path into the current process.\n This allows dynamically loading custom operators. For this, \n you should compile your operator and \n the static registration code into a shared library object, and then\n call ``paddleaudio.ops.load_library('path/to/libcustom.so')`` to load the\n shared object.\n After the library is loaded, it is added to the\n ``paddleaudio.ops.loaded_libraries`` attribute, a set that may be inspected\n for the paths of all libraries loaded using this function.\n Args:\n path (str): A path to a shared library to load.\n \"\"\"\n path = resolve_library_path(path)\n with dl_open_guard():\n # https://docs.python.org/3/library/ctypes.html?highlight=ctypes#loading-shared-libraries\n # Import the shared library into the process, thus running its\n # static (global) initialization code in order to register custom\n # operators with the JIT.\n ctypes.CDLL(path)\n self.loaded_libraries.add(path)\n\n\n_LIB_DIR = Path(__file__).parent / \"lib\"\n\n\ndef _get_lib_path(lib: str):\n suffix = \"pyd\" if os.name == \"nt\" else \"so\"\n path = _LIB_DIR / f\"{lib}.{suffix}\"\n return path\n\n\ndef _load_lib(lib: str) -> bool:\n \"\"\"Load extension module\n Note:\n In case `paddleaudio` is deployed with `pex` format, the library file\n is not in a standard location.\n In this case, we expect that `libpaddlleaudio` is available somewhere\n in the search path of dynamic loading mechanism, so that importing\n `_paddlleaudio` will have library loader find and load `libpaddlleaudio`.\n This is the reason why the function should not raising an error when the library\n file is not found.\n Returns:\n bool:\n True if the library file is found AND the library loaded without failure.\n False if the library file is not found (like in the case where paddlleaudio\n is deployed with pex format, thus the shared library file is\n in a non-standard location.).\n If the library file is found but there is an issue loading the library,\n (such as missing dependency) then this function raises the exception as-is.\n Raises:\n Exception:\n If the library file is found, but there is an issue loading the library file,\n (when underlying `ctype.DLL` throws an exception), this function will pass\n the exception as-is, instead of catching it and returning bool.\n The expected case is `OSError` thrown by `ctype.DLL` when a dynamic dependency\n is not found.\n This behavior was chosen because the expected failure case is not recoverable.\n If a dependency is missing, then users have to install it.\n \"\"\"\n path = _get_lib_path(lib)\n if not path.exists():\n warnings.warn(\"lib path is not exists:\" + str(path))\n return False\n ops.load_library(path)\n return True\n\n\n_FFMPEG_INITIALIZED = False\n\n\ndef _init_ffmpeg():\n global _FFMPEG_INITIALIZED\n if _FFMPEG_INITIALIZED:\n return\n\n if not paddleaudio._paddlleaudio.is_ffmpeg_available():\n raise RuntimeError(\n \"paddlleaudio is not compiled with FFmpeg integration. Please set USE_FFMPEG=1 when compiling paddlleaudio.\"\n )\n\n try:\n _load_lib(\"libpaddlleaudio_ffmpeg\")\n except OSError as err:\n raise ImportError(\n \"FFmpeg libraries are not found. Please install FFmpeg.\") from err\n\n import paddllespeech.audio._paddlleaudio_ffmpeg # noqa\n\n paddleaudio._paddlleaudio.ffmpeg_init()\n if paddleaudio._paddlleaudio.ffmpeg_get_log_level() > 8:\n paddleaudio._paddlleaudio.ffmpeg_set_log_level(8)\n\n _FFMPEG_INITIALIZED = True\n\n\ndef _init_extension():\n if not _mod_utils.is_module_available(\"paddleaudio._paddleaudio\"):\n warnings.warn(\n \"paddleaudio C++ extension is not available. sox_io, sox_effect, kaldi raw feature is not supported!!!\")\n return\n\n _load_lib(\"libpaddleaudio\")\n # This import is for initializing the methods registered via PyBind11\n # This has to happen after the base library is loaded\n try:\n from paddleaudio import _paddleaudio # noqa\n except Exception:\n warnings.warn(\n \"paddleaudio C++ extension is not available. sox_io, sox_effect, kaldi raw feature is not supported!!!\")\n return\n\n # Because this part is executed as part of `import torchaudio`, we ignore the\n # initialization failure.\n # If the FFmpeg integration is not properly initialized, then detailed error\n # will be raised when client code attempts to import the dedicated feature.\n try:\n _init_ffmpeg()\n except Exception:\n pass\n\n\nops = _Ops()\n\n_init_extension()\n"
},
{
"path": "audio/paddleaudio/_internal/__init__.py",
"content": ""
},
{
"path": "audio/paddleaudio/_internal/module_utils.py",
"content": "import importlib.util\nimport platform\nimport warnings\nfrom functools import wraps\nfrom typing import Optional\n\n#code is from https://github.com/pytorch/audio/blob/main/torchaudio/_internal/module_utils.py with modification.\n\n\ndef is_module_available(*modules: str) -> bool:\n r\"\"\"Returns if a top-level module with :attr:`name` exists *without**\n importing it. This is generally safer than try-catch block around a\n `import X`. It avoids third party libraries breaking assumptions of some of\n our tests, e.g., setting multiprocessing start method when imported\n (see librosa/#747, torchvision/#544).\n \"\"\"\n return all(importlib.util.find_spec(m) is not None for m in modules)\n\n\ndef requires_module(*modules: str):\n \"\"\"Decorate function to give error message if invoked without required optional modules.\n This decorator is to give better error message to users rather\n than raising ``NameError: name 'module' is not defined`` at random places.\n \"\"\"\n missing = [m for m in modules if not is_module_available(m)]\n\n if not missing:\n # fall through. If all the modules are available, no need to decorate\n def decorator(func):\n return func\n\n else:\n req = f\"module: {missing[0]}\" if len(\n missing) == 1 else f\"modules: {missing}\"\n\n def decorator(func):\n @wraps(func)\n def wrapped(*args, **kwargs):\n raise RuntimeError(\n f\"{func.__module__}.{func.__name__} requires {req}\")\n\n return wrapped\n\n return decorator\n\n\ndef deprecated(direction: str, version: Optional[str]=None):\n \"\"\"Decorator to add deprecation message\n Args:\n direction (str): Migration steps to be given to users.\n version (str or int): The version when the object will be removed\n \"\"\"\n\n def decorator(func):\n @wraps(func)\n def wrapped(*args, **kwargs):\n message = (\n f\"{func.__module__}.{func.__name__} has been deprecated \"\n f'and will be removed from {\"future\" if version is None else version} release. '\n f\"{direction}\")\n warnings.warn(message, stacklevel=2)\n return func(*args, **kwargs)\n\n return wrapped\n\n return decorator\n\n\ndef is_kaldi_available():\n try:\n from paddleaudio import _paddleaudio \n return True\n except Exception:\n return False\n\ndef requires_kaldi():\n if is_kaldi_available():\n\n def decorator(func):\n return func\n\n else:\n\n def decorator(func):\n @wraps(func)\n def wrapped(*args, **kwargs):\n raise RuntimeError(\n f\"{func.__module__}.{func.__name__} requires libpaddleaudio build with kaldi\")\n\n return wrapped\n\n return decorator\n\n\ndef _check_soundfile_importable():\n if not is_module_available(\"soundfile\"):\n return False\n try:\n import soundfile # noqa: F401\n\n return True\n except Exception:\n warnings.warn(\n \"Failed to import soundfile. 'soundfile' backend is not available.\")\n return False\n\n\n_is_soundfile_importable = _check_soundfile_importable()\n\n\ndef is_soundfile_available():\n return _is_soundfile_importable\n\n\ndef requires_soundfile():\n if is_soundfile_available():\n\n def decorator(func):\n return func\n else:\n\n def decorator(func):\n @wraps(func)\n def wrapped(*args, **kwargs):\n raise RuntimeError(\n f\"{func.__module__}.{func.__name__} requires soundfile\")\n\n return wrapped\n\n return decorator\n\n\ndef is_sox_available():\n try:\n from paddleaudio import _paddleaudio \n return True\n except Exception:\n return False\n\n\ndef requires_sox():\n if is_sox_available():\n\n def decorator(func):\n return func\n else:\n\n def decorator(func):\n @wraps(func)\n def wrapped(*args, **kwargs):\n raise RuntimeError(\n f\"{func.__module__}.{func.__name__} requires libpaddleaudio build with sox\")\n\n return wrapped\n\n return decorator\n"
},
{
"path": "audio/paddleaudio/backends/__init__.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom . import utils\nfrom .soundfile_backend import depth_convert\nfrom .soundfile_backend import normalize\nfrom .soundfile_backend import resample\nfrom .soundfile_backend import soundfile_load\nfrom .soundfile_backend import soundfile_save\nfrom .soundfile_backend import to_mono\nfrom .utils import get_audio_backend\nfrom .utils import list_audio_backends\nfrom .utils import set_audio_backend\n\nutils._init_audio_backend()\n"
},
{
"path": "audio/paddleaudio/backends/common.py",
"content": "# Token from https://github.com/pytorch/audio/blob/main/torchaudio/backend/common.py with modification.\n\n\nclass AudioInfo:\n \"\"\"return of info function.\n\n This class is used by :ref:`\"sox_io\" backend` and\n :ref:`\"soundfile\" backend with the new interface`.\n\n :ivar int sample_rate: Sample rate\n :ivar int num_frames: The number of frames\n :ivar int num_channels: The number of channels\n :ivar int bits_per_sample: The number of bits per sample. This is 0 for lossy formats,\n or when it cannot be accurately inferred.\n :ivar str encoding: Audio encoding\n The values encoding can take are one of the following:\n\n * ``PCM_S``: Signed integer linear PCM\n * ``PCM_U``: Unsigned integer linear PCM\n * ``PCM_F``: Floating point linear PCM\n * ``FLAC``: Flac, Free Lossless Audio Codec\n * ``ULAW``: Mu-law\n * ``ALAW``: A-law\n * ``MP3`` : MP3, MPEG-1 Audio Layer III\n * ``VORBIS``: OGG Vorbis\n * ``AMR_WB``: Adaptive Multi-Rate\n * ``AMR_NB``: Adaptive Multi-Rate Wideband\n * ``OPUS``: Opus\n * ``HTK``: Single channel 16-bit PCM\n * ``UNKNOWN`` : None of above\n \"\"\"\n\n def __init__(\n self,\n sample_rate: int,\n num_frames: int,\n num_channels: int,\n bits_per_sample: int,\n encoding: str, ):\n self.sample_rate = sample_rate\n self.num_frames = num_frames\n self.num_channels = num_channels\n self.bits_per_sample = bits_per_sample\n self.encoding = encoding\n\n def __str__(self):\n return (f\"AudioMetaData(\"\n f\"sample_rate={self.sample_rate}, \"\n f\"num_frames={self.num_frames}, \"\n f\"num_channels={self.num_channels}, \"\n f\"bits_per_sample={self.bits_per_sample}, \"\n f\"encoding={self.encoding}\"\n f\")\")\n"
},
{
"path": "audio/paddleaudio/backends/no_backend.py",
"content": "from pathlib import Path\nfrom typing import Callable\nfrom typing import Optional\nfrom typing import Tuple\nfrom typing import Union\n\nfrom paddle import Tensor\n\n#code is from: https://github.com/pytorch/audio/blob/main/torchaudio/backend/no_backend.py\n\n\ndef load(\n filepath: Union[str, Path],\n out: Optional[Tensor]=None,\n normalization: Union[bool, float, Callable]=True,\n channels_first: bool=True,\n num_frames: int=0,\n offset: int=0,\n filetype: Optional[str]=None, ) -> Tuple[Tensor, int]:\n raise RuntimeError(\"No audio I/O backend is available.\")\n\n\ndef save(filepath: str,\n src: Tensor,\n sample_rate: int,\n precision: int=16,\n channels_first: bool=True) -> None:\n raise RuntimeError(\"No audio I/O backend is available.\")\n\n\ndef info(filepath: str) -> None:\n raise RuntimeError(\"No audio I/O backend is available.\")\n"
},
{
"path": "audio/paddleaudio/backends/soundfile_backend.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport os\nimport warnings\nfrom typing import Optional\nfrom typing import Tuple\n\nimport numpy as np\nimport paddle\nimport resampy\nimport soundfile\nfrom scipy.io import wavfile\n\nfrom ..utils import depth_convert\nfrom ..utils import ParameterError\nfrom .common import AudioInfo\n\n__all__ = [\n 'resample',\n 'to_mono',\n 'normalize',\n 'save',\n 'soundfile_save',\n 'load',\n 'soundfile_load',\n 'info',\n]\nNORMALMIZE_TYPES = ['linear', 'gaussian']\nMERGE_TYPES = ['ch0', 'ch1', 'random', 'average']\nRESAMPLE_MODES = ['kaiser_best', 'kaiser_fast']\nEPS = 1e-8\n\n\ndef resample(y: np.ndarray,\n src_sr: int,\n target_sr: int,\n mode: str='kaiser_fast') -> np.ndarray:\n \"\"\"Audio resampling.\n\n Args:\n y (np.ndarray): Input waveform array in 1D or 2D.\n src_sr (int): Source sample rate.\n target_sr (int): Target sample rate.\n mode (str, optional): The resampling filter to use. Defaults to 'kaiser_fast'.\n\n Returns:\n np.ndarray: `y` resampled to `target_sr`\n \"\"\"\n\n if mode == 'kaiser_best':\n warnings.warn(\n f'Using resampy in kaiser_best to {src_sr}=>{target_sr}. This function is pretty slow, \\\n we recommend the mode kaiser_fast in large scale audio training')\n\n if not isinstance(y, np.ndarray):\n raise ParameterError(\n 'Only support numpy np.ndarray, but received y in {type(y)}')\n\n if mode not in RESAMPLE_MODES:\n raise ParameterError(f'resample mode must in {RESAMPLE_MODES}')\n\n return resampy.resample(y, src_sr, target_sr, filter=mode)\n\n\ndef to_mono(y: np.ndarray, merge_type: str='average') -> np.ndarray:\n \"\"\"Convert sterior audio to mono.\n\n Args:\n y (np.ndarray): Input waveform array in 1D or 2D.\n merge_type (str, optional): Merge type to generate mono waveform. Defaults to 'average'.\n\n Returns:\n np.ndarray: `y` with mono channel.\n \"\"\"\n\n if merge_type not in MERGE_TYPES:\n raise ParameterError(\n f'Unsupported merge type {merge_type}, available types are {MERGE_TYPES}'\n )\n if y.ndim > 2:\n raise ParameterError(\n f'Unsupported audio array, y.ndim > 2, the shape is {y.shape}')\n if y.ndim == 1: # nothing to merge\n return y\n\n if merge_type == 'ch0':\n return y[0]\n if merge_type == 'ch1':\n return y[1]\n if merge_type == 'random':\n return y[np.random.randint(0, 2)]\n\n # need to do averaging according to dtype\n\n if y.dtype == 'float32':\n y_out = (y[0] + y[1]) * 0.5\n elif y.dtype == 'int16':\n y_out = y.astype('int32')\n y_out = (y_out[0] + y_out[1]) // 2\n y_out = np.clip(y_out, np.iinfo(y.dtype).min,\n np.iinfo(y.dtype).max).astype(y.dtype)\n\n elif y.dtype == 'int8':\n y_out = y.astype('int16')\n y_out = (y_out[0] + y_out[1]) // 2\n y_out = np.clip(y_out, np.iinfo(y.dtype).min,\n np.iinfo(y.dtype).max).astype(y.dtype)\n else:\n raise ParameterError(f'Unsupported dtype: {y.dtype}')\n return y_out\n\n\ndef soundfile_load_(file: os.PathLike,\n offset: Optional[float]=None,\n dtype: str='int16',\n duration: Optional[int]=None) -> Tuple[np.ndarray, int]:\n \"\"\"Load audio using soundfile library. This function load audio file using libsndfile.\n\n Args:\n file (os.PathLike): File of waveform.\n offset (Optional[float], optional): Offset to the start of waveform. Defaults to None.\n dtype (str, optional): Data type of waveform. Defaults to 'int16'.\n duration (Optional[int], optional): Duration of waveform to read. Defaults to None.\n\n Returns:\n Tuple[np.ndarray, int]: Waveform in ndarray and its samplerate.\n \"\"\"\n with soundfile.SoundFile(file) as sf_desc:\n sr_native = sf_desc.samplerate\n if offset:\n sf_desc.seek(int(offset * sr_native))\n if duration is not None:\n frame_duration = int(duration * sr_native)\n else:\n frame_duration = -1\n y = sf_desc.read(frames=frame_duration, dtype=dtype, always_2d=False).T\n\n return y, sf_desc.samplerate\n\n\ndef normalize(y: np.ndarray, norm_type: str='linear',\n mul_factor: float=1.0) -> np.ndarray:\n \"\"\"Normalize an input audio with additional multiplier.\n\n Args:\n y (np.ndarray): Input waveform array in 1D or 2D.\n norm_type (str, optional): Type of normalization. Defaults to 'linear'.\n mul_factor (float, optional): Scaling factor. Defaults to 1.0.\n\n Returns:\n np.ndarray: `y` after normalization.\n \"\"\"\n\n if norm_type == 'linear':\n amax = np.max(np.abs(y))\n factor = 1.0 / (amax + EPS)\n y = y * factor * mul_factor\n elif norm_type == 'gaussian':\n amean = np.mean(y)\n astd = np.std(y)\n astd = max(astd, EPS)\n y = mul_factor * (y - amean) / astd\n else:\n raise NotImplementedError(f'norm_type should be in {NORMALMIZE_TYPES}')\n\n return y\n\n\ndef soundfile_save(y: np.ndarray, sr: int, file: os.PathLike) -> None:\n \"\"\"Save audio file to disk. This function saves audio to disk using scipy.io.wavfile, with additional step to convert input waveform to int16.\n\n Args:\n y (np.ndarray): Input waveform array in 1D or 2D.\n sr (int): Sample rate.\n file (os.PathLike): Path of audio file to save.\n \"\"\"\n if not file.endswith('.wav'):\n raise ParameterError(\n f'only .wav file supported, but dst file name is: {file}')\n\n if sr <= 0:\n raise ParameterError(\n f'Sample rate should be larger than 0, received sr = {sr}')\n\n if y.dtype not in ['int16', 'int8']:\n warnings.warn(\n f'input data type is {y.dtype}, will convert data to int16 format before saving'\n )\n y_out = depth_convert(y, 'int16')\n else:\n y_out = y\n\n wavfile.write(file, sr, y_out)\n\n\ndef soundfile_load(\n file: os.PathLike,\n sr: Optional[int]=None,\n mono: bool=True,\n merge_type: str='average', # ch0,ch1,random,average\n normal: bool=True,\n norm_type: str='linear',\n norm_mul_factor: float=1.0,\n offset: float=0.0,\n duration: Optional[int]=None,\n dtype: str='float32',\n resample_mode: str='kaiser_fast') -> Tuple[np.ndarray, int]:\n \"\"\"Load audio file from disk. This function loads audio from disk using using audio backend.\n\n Args:\n file (os.PathLike): Path of audio file to load.\n sr (Optional[int], optional): Sample rate of loaded waveform. Defaults to None.\n mono (bool, optional): Return waveform with mono channel. Defaults to True.\n merge_type (str, optional): Merge type of multi-channels waveform. Defaults to 'average'.\n normal (bool, optional): Waveform normalization. Defaults to True.\n norm_type (str, optional): Type of normalization. Defaults to 'linear'.\n norm_mul_factor (float, optional): Scaling factor. Defaults to 1.0.\n offset (float, optional): Offset to the start of waveform. Defaults to 0.0.\n duration (Optional[int], optional): Duration of waveform to read. Defaults to None.\n dtype (str, optional): Data type of waveform. Defaults to 'float32'.\n resample_mode (str, optional): The resampling filter to use. Defaults to 'kaiser_fast'.\n\n Returns:\n Tuple[np.ndarray, int]: Waveform in ndarray and its samplerate.\n \"\"\"\n\n y, r = soundfile_load_(file, offset=offset, dtype=dtype, duration=duration)\n\n if not ((y.ndim == 1 and len(y) > 0) or (y.ndim == 2 and len(y[0]) > 0)):\n raise ParameterError(f'audio file {file} looks empty')\n\n if mono:\n y = to_mono(y, merge_type)\n\n if sr is not None and sr != r:\n y = resample(y, r, sr, mode=resample_mode)\n r = sr\n\n if normal:\n y = normalize(y, norm_type, norm_mul_factor)\n elif dtype in ['int8', 'int16']:\n # still need to do normalization, before depth conversion\n y = normalize(y, 'linear', 1.0)\n\n y = depth_convert(y, dtype)\n return y, r\n\n\n#The code below is taken from: https://github.com/pytorch/audio/blob/main/torchaudio/backend/soundfile_backend.py, with some modifications.\n\n\ndef _get_subtype_for_wav(dtype: paddle.dtype,\n encoding: str,\n bits_per_sample: int):\n if not encoding:\n if not bits_per_sample:\n subtype = {\n paddle.uint8: \"PCM_U8\",\n paddle.int16: \"PCM_16\",\n paddle.int32: \"PCM_32\",\n paddle.float32: \"FLOAT\",\n paddle.float64: \"DOUBLE\",\n }.get(dtype)\n if not subtype:\n raise ValueError(f\"Unsupported dtype for wav: {dtype}\")\n return subtype\n if bits_per_sample == 8:\n return \"PCM_U8\"\n return f\"PCM_{bits_per_sample}\"\n if encoding == \"PCM_S\":\n if not bits_per_sample:\n return \"PCM_32\"\n if bits_per_sample == 8:\n raise ValueError(\"wav does not support 8-bit signed PCM encoding.\")\n return f\"PCM_{bits_per_sample}\"\n if encoding == \"PCM_U\":\n if bits_per_sample in (None, 8):\n return \"PCM_U8\"\n raise ValueError(\"wav only supports 8-bit unsigned PCM encoding.\")\n if encoding == \"PCM_F\":\n if bits_per_sample in (None, 32):\n return \"FLOAT\"\n if bits_per_sample == 64:\n return \"DOUBLE\"\n raise ValueError(\"wav only supports 32/64-bit float PCM encoding.\")\n if encoding == \"ULAW\":\n if bits_per_sample in (None, 8):\n return \"ULAW\"\n raise ValueError(\"wav only supports 8-bit mu-law encoding.\")\n if encoding == \"ALAW\":\n if bits_per_sample in (None, 8):\n return \"ALAW\"\n raise ValueError(\"wav only supports 8-bit a-law encoding.\")\n raise ValueError(f\"wav does not support {encoding}.\")\n\n\ndef _get_subtype_for_sphere(encoding: str, bits_per_sample: int):\n if encoding in (None, \"PCM_S\"):\n return f\"PCM_{bits_per_sample}\" if bits_per_sample else \"PCM_32\"\n if encoding in (\"PCM_U\", \"PCM_F\"):\n raise ValueError(f\"sph does not support {encoding} encoding.\")\n if encoding == \"ULAW\":\n if bits_per_sample in (None, 8):\n return \"ULAW\"\n raise ValueError(\"sph only supports 8-bit for mu-law encoding.\")\n if encoding == \"ALAW\":\n return \"ALAW\"\n raise ValueError(f\"sph does not support {encoding}.\")\n\n\ndef _get_subtype(dtype: paddle.dtype,\n format: str,\n encoding: str,\n bits_per_sample: int):\n if format == \"wav\":\n return _get_subtype_for_wav(dtype, encoding, bits_per_sample)\n if format == \"flac\":\n if encoding:\n raise ValueError(\"flac does not support encoding.\")\n if not bits_per_sample:\n return \"PCM_16\"\n if bits_per_sample > 24:\n raise ValueError(\"flac does not support bits_per_sample > 24.\")\n return \"PCM_S8\" if bits_per_sample == 8 else f\"PCM_{bits_per_sample}\"\n if format in (\"ogg\", \"vorbis\"):\n if encoding or bits_per_sample:\n raise ValueError(\n \"ogg/vorbis does not support encoding/bits_per_sample.\")\n return \"VORBIS\"\n if format == \"sph\":\n return _get_subtype_for_sphere(encoding, bits_per_sample)\n if format in (\"nis\", \"nist\"):\n return \"PCM_16\"\n raise ValueError(f\"Unsupported format: {format}\")\n\n\ndef save(\n filepath: str,\n src: paddle.Tensor,\n sample_rate: int,\n channels_first: bool=True,\n compression: Optional[float]=None,\n format: Optional[str]=None,\n encoding: Optional[str]=None,\n bits_per_sample: Optional[int]=None, ):\n \"\"\"Save audio data to file.\n\n Note:\n The formats this function can handle depend on the soundfile installation.\n This function is tested on the following formats;\n\n * WAV\n\n * 32-bit floating-point\n * 32-bit signed integer\n * 16-bit signed integer\n * 8-bit unsigned integer\n\n * FLAC\n * OGG/VORBIS\n * SPHERE\n\n Note:\n ``filepath`` argument is intentionally annotated as ``str`` only, even though it accepts\n ``pathlib.Path`` object as well. This is for the consistency with ``\"sox_io\"`` backend,\n\n Args:\n filepath (str or pathlib.Path): Path to audio file.\n src (paddle.Tensor): Audio data to save. must be 2D tensor.\n sample_rate (int): sampling rate\n channels_first (bool, optional): If ``True``, the given tensor is interpreted as `[channel, time]`,\n otherwise `[time, channel]`.\n compression (float of None, optional): Not used.\n It is here only for interface compatibility reason with \"sox_io\" backend.\n format (str or None, optional): Override the audio format.\n When ``filepath`` argument is path-like object, audio format is\n inferred from file extension. If the file extension is missing or\n different, you can specify the correct format with this argument.\n\n When ``filepath`` argument is file-like object,\n this argument is required.\n\n Valid values are ``\"wav\"``, ``\"ogg\"``, ``\"vorbis\"``,\n ``\"flac\"`` and ``\"sph\"``.\n encoding (str or None, optional): Changes the encoding for supported formats.\n This argument is effective only for supported formats, such as\n ``\"wav\"``, ``\"\"flac\"`` and ``\"sph\"``. Valid values are:\n\n - ``\"PCM_S\"`` (signed integer Linear PCM)\n - ``\"PCM_U\"`` (unsigned integer Linear PCM)\n - ``\"PCM_F\"`` (floating point PCM)\n - ``\"ULAW\"`` (mu-law)\n - ``\"ALAW\"`` (a-law)\n\n bits_per_sample (int or None, optional): Changes the bit depth for the\n supported formats.\n When ``format`` is one of ``\"wav\"``, ``\"flac\"`` or ``\"sph\"``,\n you can change the bit depth.\n Valid values are ``8``, ``16``, ``24``, ``32`` and ``64``.\n\n Supported formats/encodings/bit depth/compression are:\n\n ``\"wav\"``\n - 32-bit floating-point PCM\n - 32-bit signed integer PCM\n - 24-bit signed integer PCM\n - 16-bit signed integer PCM\n - 8-bit unsigned integer PCM\n - 8-bit mu-law\n - 8-bit a-law\n\n Note:\n Default encoding/bit depth is determined by the dtype of\n the input Tensor.\n\n ``\"flac\"``\n - 8-bit\n - 16-bit (default)\n - 24-bit\n\n ``\"ogg\"``, ``\"vorbis\"``\n - Doesn't accept changing configuration.\n\n ``\"sph\"``\n - 8-bit signed integer PCM\n - 16-bit signed integer PCM\n - 24-bit signed integer PCM\n - 32-bit signed integer PCM (default)\n - 8-bit mu-law\n - 8-bit a-law\n - 16-bit a-law\n - 24-bit a-law\n - 32-bit a-law\n\n \"\"\"\n if src.ndim != 2:\n raise ValueError(f\"Expected 2D Tensor, got {src.ndim}D.\")\n if compression is not None:\n warnings.warn(\n '`save` function of \"soundfile\" backend does not support \"compression\" parameter. '\n \"The argument is silently ignored.\")\n if hasattr(filepath, \"write\"):\n if format is None:\n raise RuntimeError(\n \"`format` is required when saving to file object.\")\n ext = format.lower()\n else:\n ext = str(filepath).split(\".\")[-1].lower()\n\n if bits_per_sample not in (None, 8, 16, 24, 32, 64):\n raise ValueError(\"Invalid bits_per_sample.\")\n if bits_per_sample == 24:\n warnings.warn(\n \"Saving audio with 24 bits per sample might warp samples near -1. \"\n \"Using 16 bits per sample might be able to avoid this.\")\n subtype = _get_subtype(src.dtype, ext, encoding, bits_per_sample)\n\n # sph is a extension used in TED-LIUM but soundfile does not recognize it as NIST format,\n # so we extend the extensions manually here\n if ext in [\"nis\", \"nist\", \"sph\"] and format is None:\n format = \"NIST\"\n\n if channels_first:\n src = src.t()\n\n soundfile.write(\n file=filepath,\n data=src,\n samplerate=sample_rate,\n subtype=subtype,\n format=format)\n\n\n_SUBTYPE2DTYPE = {\n \"PCM_S8\": \"int8\",\n \"PCM_U8\": \"uint8\",\n \"PCM_16\": \"int16\",\n \"PCM_32\": \"int32\",\n \"FLOAT\": \"float32\",\n \"DOUBLE\": \"float64\",\n}\n\n\ndef load(\n filepath: str,\n frame_offset: int=0,\n num_frames: int=-1,\n normalize: bool=True,\n channels_first: bool=True,\n format: Optional[str]=None, ) -> Tuple[paddle.Tensor, int]:\n \"\"\"Load audio data from file.\n\n Note:\n The formats this function can handle depend on the soundfile installation.\n This function is tested on the following formats;\n\n * WAV\n\n * 32-bit floating-point\n * 32-bit signed integer\n * 16-bit signed integer\n * 8-bit unsigned integer\n\n * FLAC\n * OGG/VORBIS\n * SPHERE\n\n By default (``normalize=True``, ``channels_first=True``), this function returns Tensor with\n ``float32`` dtype and the shape of `[channel, time]`.\n The samples are normalized to fit in the range of ``[-1.0, 1.0]``.\n\n When the input format is WAV with integer type, such as 32-bit signed integer, 16-bit\n signed integer and 8-bit unsigned integer (24-bit signed integer is not supported),\n by providing ``normalize=False``, this function can return integer Tensor, where the samples\n are expressed within the whole range of the corresponding dtype, that is, ``int32`` tensor\n for 32-bit signed PCM, ``int16`` for 16-bit signed PCM and ``uint8`` for 8-bit unsigned PCM.\n\n ``normalize`` parameter has no effect on 32-bit floating-point WAV and other formats, such as\n ``flac`` and ``mp3``.\n For these formats, this function always returns ``float32`` Tensor with values normalized to\n ``[-1.0, 1.0]``.\n\n Note:\n ``filepath`` argument is intentionally annotated as ``str`` only, even though it accepts\n ``pathlib.Path`` object as well. This is for the consistency with ``\"sox_io\"`` backend.\n\n Args:\n filepath (path-like object or file-like object):\n Source of audio data.\n frame_offset (int, optional):\n Number of frames to skip before start reading data.\n num_frames (int, optional):\n Maximum number of frames to read. ``-1`` reads all the remaining samples,\n starting from ``frame_offset``.\n This function may return the less number of frames if there is not enough\n frames in the given file.\n normalize (bool, optional):\n When ``True``, this function always return ``float32``, and sample values are\n normalized to ``[-1.0, 1.0]``.\n If input file is integer WAV, giving ``False`` will change the resulting Tensor type to\n integer type.\n This argument has no effect for formats other than integer WAV type.\n channels_first (bool, optional):\n When True, the returned Tensor has dimension `[channel, time]`.\n Otherwise, the returned Tensor's dimension is `[time, channel]`.\n format (str or None, optional):\n Not used. PySoundFile does not accept format hint.\n\n Returns:\n (paddle.Tensor, int): Resulting Tensor and sample rate.\n If the input file has integer wav format and normalization is off, then it has\n integer type, else ``float32`` type. If ``channels_first=True``, it has\n `[channel, time]` else `[time, channel]`.\n \"\"\"\n with soundfile.SoundFile(filepath, \"r\") as file_:\n if file_.format != \"WAV\" or normalize:\n dtype = \"float32\"\n elif file_.subtype not in _SUBTYPE2DTYPE:\n raise ValueError(f\"Unsupported subtype: {file_.subtype}\")\n else:\n dtype = _SUBTYPE2DTYPE[file_.subtype]\n\n frames = file_._prepare_read(frame_offset, None, num_frames)\n waveform = file_.read(frames, dtype, always_2d=True)\n sample_rate = file_.samplerate\n\n waveform = paddle.to_tensor(waveform)\n if channels_first:\n waveform = paddle.transpose(waveform, perm=[1, 0])\n return waveform, sample_rate\n\n\n# Mapping from soundfile subtype to number of bits per sample.\n# This is mostly heuristical and the value is set to 0 when it is irrelevant\n# (lossy formats) or when it can't be inferred.\n# For ADPCM (and G72X) subtypes, it's hard to infer the bit depth because it's not part of the standard:\n# According to https://en.wikipedia.org/wiki/Adaptive_differential_pulse-code_modulation#In_telephony,\n# the default seems to be 8 bits but it can be compressed further to 4 bits.\n# The dict is inspired from\n# https://github.com/bastibe/python-soundfile/blob/744efb4b01abc72498a96b09115b42a4cabd85e4/soundfile.py#L66-L94\n_SUBTYPE_TO_BITS_PER_SAMPLE = {\n \"PCM_S8\": 8, # Signed 8 bit data\n \"PCM_16\": 16, # Signed 16 bit data\n \"PCM_24\": 24, # Signed 24 bit data\n \"PCM_32\": 32, # Signed 32 bit data\n \"PCM_U8\": 8, # Unsigned 8 bit data (WAV and RAW only)\n \"FLOAT\": 32, # 32 bit float data\n \"DOUBLE\": 64, # 64 bit float data\n \"ULAW\": 8, # U-Law encoded. See https://en.wikipedia.org/wiki/G.711#Types\n \"ALAW\": 8, # A-Law encoded. See https://en.wikipedia.org/wiki/G.711#Types\n \"IMA_ADPCM\": 0, # IMA ADPCM.\n \"MS_ADPCM\": 0, # Microsoft ADPCM.\n \"GSM610\":\n 0, # GSM 6.10 encoding. (Wikipedia says 1.625 bit depth?? https://en.wikipedia.org/wiki/Full_Rate)\n \"VOX_ADPCM\": 0, # OKI / Dialogix ADPCM\n \"G721_32\": 0, # 32kbs G721 ADPCM encoding.\n \"G723_24\": 0, # 24kbs G723 ADPCM encoding.\n \"G723_40\": 0, # 40kbs G723 ADPCM encoding.\n \"DWVW_12\": 12, # 12 bit Delta Width Variable Word encoding.\n \"DWVW_16\": 16, # 16 bit Delta Width Variable Word encoding.\n \"DWVW_24\": 24, # 24 bit Delta Width Variable Word encoding.\n \"DWVW_N\": 0, # N bit Delta Width Variable Word encoding.\n \"DPCM_8\": 8, # 8 bit differential PCM (XI only)\n \"DPCM_16\": 16, # 16 bit differential PCM (XI only)\n \"VORBIS\": 0, # Xiph Vorbis encoding. (lossy)\n \"ALAC_16\": 16, # Apple Lossless Audio Codec (16 bit).\n \"ALAC_20\": 20, # Apple Lossless Audio Codec (20 bit).\n \"ALAC_24\": 24, # Apple Lossless Audio Codec (24 bit).\n \"ALAC_32\": 32, # Apple Lossless Audio Codec (32 bit).\n}\n\n\ndef _get_bit_depth(subtype):\n if subtype not in _SUBTYPE_TO_BITS_PER_SAMPLE:\n warnings.warn(\n f\"The {subtype} subtype is unknown to PaddleAudio. As a result, the bits_per_sample \"\n \"attribute will be set to 0. If you are seeing this warning, please \"\n \"report by opening an issue on github (after checking for existing/closed ones). \"\n \"You may otherwise ignore this warning.\")\n return _SUBTYPE_TO_BITS_PER_SAMPLE.get(subtype, 0)\n\n\n_SUBTYPE_TO_ENCODING = {\n \"PCM_S8\": \"PCM_S\",\n \"PCM_16\": \"PCM_S\",\n \"PCM_24\": \"PCM_S\",\n \"PCM_32\": \"PCM_S\",\n \"PCM_U8\": \"PCM_U\",\n \"FLOAT\": \"PCM_F\",\n \"DOUBLE\": \"PCM_F\",\n \"ULAW\": \"ULAW\",\n \"ALAW\": \"ALAW\",\n \"VORBIS\": \"VORBIS\",\n}\n\n\ndef _get_encoding(format: str, subtype: str):\n if format == \"FLAC\":\n return \"FLAC\"\n return _SUBTYPE_TO_ENCODING.get(subtype, \"UNKNOWN\")\n\n\ndef info(filepath: str, format: Optional[str]=None) -> AudioInfo:\n \"\"\"Get signal information of an audio file.\n\n Note:\n ``filepath`` argument is intentionally annotated as ``str`` only, even though it accepts\n ``pathlib.Path`` object as well. This is for the consistency with ``\"sox_io\"`` backend,\n\n Args:\n filepath (path-like object or file-like object):\n Source of audio data.\n format (str or None, optional):\n Not used. PySoundFile does not accept format hint.\n\n Returns:\n AudioInfo: meta data of the given audio.\n\n \"\"\"\n sinfo = soundfile.info(filepath)\n return AudioInfo(\n sinfo.samplerate,\n sinfo.frames,\n sinfo.channels,\n bits_per_sample=_get_bit_depth(sinfo.subtype),\n encoding=_get_encoding(sinfo.format, sinfo.subtype), )\n"
},
{
"path": "audio/paddleaudio/backends/sox_io_backend.py",
"content": "import os\nfrom typing import Optional\nfrom typing import Tuple\n\nimport paddle\nimport paddleaudio\nfrom paddle import Tensor\nfrom paddleaudio._internal import module_utils as _mod_utils\n\nfrom .common import AudioInfo\n\n#https://github.com/pytorch/audio/blob/main/torchaudio/backend/sox_io_backend.py\n\n\ndef _fail_info(filepath: str, format: Optional[str]) -> AudioInfo:\n raise RuntimeError(\"Failed to fetch metadata from {}\".format(filepath))\n\n\ndef _fail_info_fileobj(fileobj, format: Optional[str]) -> AudioInfo:\n raise RuntimeError(\"Failed to fetch metadata from {}\".format(fileobj))\n\n\n# Note: need to comply TorchScript syntax -- need annotation and no f-string\ndef _fail_load(\n filepath: str,\n frame_offset: int=0,\n num_frames: int=-1,\n normalize: bool=True,\n channels_first: bool=True,\n format: Optional[str]=None, ) -> Tuple[Tensor, int]:\n raise RuntimeError(\"Failed to load audio from {}\".format(filepath))\n\n\ndef _fail_load_fileobj(fileobj, *args, **kwargs):\n raise RuntimeError(f\"Failed to load audio from {fileobj}\")\n\n\n_fallback_info = _fail_info\n_fallback_info_fileobj = _fail_info_fileobj\n_fallback_load = _fail_load\n_fallback_load_filebj = _fail_load_fileobj\n\n\n@_mod_utils.requires_sox()\ndef load(\n filepath: str,\n frame_offset: int=0,\n num_frames: int=-1,\n normalize: bool=True,\n channels_first: bool=True,\n format: Optional[str]=None, ) -> Tuple[Tensor, int]:\n if hasattr(filepath, \"read\"):\n ret = paddleaudio._paddleaudio.load_audio_fileobj(\n filepath, frame_offset, num_frames, normalize, channels_first,\n format)\n if ret is not None:\n audio_tensor = paddle.to_tensor(ret[0])\n return (audio_tensor, ret[1])\n return _fallback_load_fileobj(filepath, frame_offset, num_frames,\n normalize, channels_first, format)\n filepath = os.fspath(filepath)\n ret = paddleaudio._paddleaudio.sox_io_load_audio_file(\n filepath, frame_offset, num_frames, normalize, channels_first, format)\n if ret is not None:\n audio_tensor = paddle.to_tensor(ret[0])\n return (audio_tensor, ret[1])\n return _fallback_load(filepath, frame_offset, num_frames, normalize,\n channels_first, format)\n\n\n@_mod_utils.requires_sox()\ndef save(\n filepath: str,\n src: Tensor,\n sample_rate: int,\n channels_first: bool=True,\n compression: Optional[float]=None,\n format: Optional[str]=None,\n encoding: Optional[str]=None,\n bits_per_sample: Optional[int]=None, ):\n src_arr = src.numpy()\n if hasattr(filepath, \"write\"):\n paddleaudio._paddleaudio.save_audio_fileobj(\n filepath, src_arr, sample_rate, channels_first, compression, format,\n encoding, bits_per_sample)\n return\n filepath = os.fspath(filepath)\n paddleaudio._paddleaudio.sox_io_save_audio_file(\n filepath, src_arr, sample_rate, channels_first, compression, format,\n encoding, bits_per_sample)\n\n\n@_mod_utils.requires_sox()\ndef info(\n filepath: str,\n format: Optional[str]=None, ) -> AudioInfo:\n if hasattr(filepath, \"read\"):\n sinfo = paddleaudio._paddleaudio.get_info_fileobj(filepath, format)\n if sinfo is not None:\n return AudioInfo(*sinfo)\n return _fallback_info_fileobj(filepath, format)\n filepath = os.fspath(filepath)\n sinfo = paddleaudio._paddleaudio.get_info_file(filepath, format)\n if sinfo is not None:\n return AudioInfo(*sinfo)\n return _fallback_info(filepath, format)\n"
},
{
"path": "audio/paddleaudio/backends/utils.py",
"content": "\"\"\"Defines utilities for switching audio backends\"\"\"\n#code is from: https://github.com/pytorch/audio/blob/main/torchaudio/backend/utils.py\nimport warnings\nfrom typing import List\nfrom typing import Optional\n\nimport paddleaudio\nfrom paddleaudio._internal import module_utils as _mod_utils\n\nfrom . import no_backend\nfrom . import soundfile_backend\nfrom . import sox_io_backend\n\n__all__ = [\n \"list_audio_backends\",\n \"get_audio_backend\",\n \"set_audio_backend\",\n]\n\n\ndef list_audio_backends() -> List[str]:\n \"\"\"List available backends\n\n Returns:\n List[str]: The list of available backends.\n \"\"\"\n backends = []\n if _mod_utils.is_module_available(\"soundfile\"):\n backends.append(\"soundfile\")\n if _mod_utils.is_sox_available():\n backends.append(\"sox_io\")\n return backends\n\n\ndef set_audio_backend(backend: Optional[str]):\n \"\"\"Set the backend for I/O operation\n\n Args:\n backend (str or None): Name of the backend.\n One of ``\"sox_io\"`` or ``\"soundfile\"`` based on availability\n of the system. If ``None`` is provided the current backend is unassigned.\n \"\"\"\n if backend is not None and backend not in list_audio_backends():\n raise RuntimeError(f'Backend \"{backend}\" is not one of '\n f\"available backends: {list_audio_backends()}.\")\n\n if backend is None:\n module = no_backend\n elif backend == \"sox_io\":\n module = sox_io_backend\n elif backend == \"soundfile\":\n module = soundfile_backend\n else:\n raise NotImplementedError(f'Unexpected backend \"{backend}\"')\n\n for func in [\"save\", \"load\", \"info\"]:\n setattr(paddleaudio, func, getattr(module, func))\n\n\ndef _init_audio_backend():\n backends = list_audio_backends()\n if \"soundfile\" in backends:\n set_audio_backend(\"soundfile\")\n elif \"sox_io\" in backends:\n set_audio_backend(\"sox_io\")\n else:\n warnings.warn(\"No audio backend is available.\")\n set_audio_backend(None)\n\n\ndef get_audio_backend() -> Optional[str]:\n \"\"\"Get the name of the current backend\n\n Returns:\n Optional[str]: The name of the current backend or ``None`` if no backend is assigned.\n \"\"\"\n if paddleaudio.load == no_backend.load:\n return None\n if paddleaudio.load == sox_io_backend.load:\n return \"sox_io\"\n if paddleaudio.load == soundfile_backend.load:\n return \"soundfile\"\n raise ValueError(\"Unknown backend.\")\n"
},
{
"path": "audio/paddleaudio/compliance/__init__.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom . import kaldi\nfrom . import librosa\n"
},
{
"path": "audio/paddleaudio/compliance/kaldi.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n# Modified from torchaudio(https://github.com/pytorch/audio)\nimport math\nfrom typing import Tuple\n\nimport paddle\nfrom paddle import Tensor\n\nfrom ..functional import create_dct\nfrom ..functional.window import get_window\n\n__all__ = [\n 'spectrogram',\n 'fbank',\n 'mfcc',\n]\n\n# window types\nHANNING = 'hann'\nHAMMING = 'hamming'\nPOVEY = 'povey'\nRECTANGULAR = 'rect'\nBLACKMAN = 'blackman'\n\n\ndef _get_epsilon(dtype):\n return paddle.to_tensor(1e-07, dtype=dtype)\n\n\ndef _next_power_of_2(x: int) -> int:\n return 1 if x == 0 else 2**(x - 1).bit_length()\n\n\ndef _get_strided(waveform: Tensor,\n window_size: int,\n window_shift: int,\n snip_edges: bool) -> Tensor:\n assert waveform.dim() == 1\n num_samples = waveform.shape[0]\n\n if snip_edges:\n if num_samples < window_size:\n return paddle.empty((0, 0), dtype=waveform.dtype)\n else:\n m = 1 + (num_samples - window_size) // window_shift\n else:\n reversed_waveform = paddle.flip(waveform, [0])\n m = (num_samples + (window_shift // 2)) // window_shift\n pad = window_size // 2 - window_shift // 2\n pad_right = reversed_waveform\n if pad > 0:\n pad_left = reversed_waveform[-pad:]\n waveform = paddle.concat((pad_left, waveform, pad_right), axis=0)\n else:\n waveform = paddle.concat((waveform[-pad:], pad_right), axis=0)\n\n return paddle.signal.frame(waveform, window_size, window_shift)[:, :m].T\n\n\ndef _feature_window_function(\n window_type: str,\n window_size: int,\n blackman_coeff: float,\n dtype: int, ) -> Tensor:\n if window_type == \"hann\":\n return get_window('hann', window_size, fftbins=False, dtype=dtype)\n elif window_type == \"hamming\":\n return get_window('hamming', window_size, fftbins=False, dtype=dtype)\n elif window_type == \"povey\":\n return get_window(\n 'hann', window_size, fftbins=False, dtype=dtype).pow(0.85)\n elif window_type == \"rect\":\n return paddle.ones([window_size], dtype=dtype)\n elif window_type == \"blackman\":\n a = 2 * math.pi / (window_size - 1)\n window_function = paddle.arange(window_size, dtype=dtype)\n return (blackman_coeff - 0.5 * paddle.cos(a * window_function) +\n (0.5 - blackman_coeff) * paddle.cos(2 * a * window_function)\n ).astype(dtype)\n else:\n raise Exception('Invalid window type ' + window_type)\n\n\ndef _get_log_energy(strided_input: Tensor, epsilon: Tensor,\n energy_floor: float) -> Tensor:\n log_energy = paddle.maximum(strided_input.pow(2).sum(1), epsilon).log()\n if energy_floor == 0.0:\n return log_energy\n return paddle.maximum(\n log_energy,\n paddle.to_tensor(math.log(energy_floor), dtype=strided_input.dtype))\n\n\ndef _get_waveform_and_window_properties(\n waveform: Tensor,\n channel: int,\n sr: int,\n frame_shift: float,\n frame_length: float,\n round_to_power_of_two: bool,\n preemphasis_coefficient: float) -> Tuple[Tensor, int, int, int]:\n channel = max(channel, 0)\n assert channel < waveform.shape[0], (\n 'Invalid channel {} for size {}'.format(channel, waveform.shape[0]))\n waveform = waveform[channel, :] # size (n)\n window_shift = int(\n sr * frame_shift *\n 0.001) # pass frame_shift and frame_length in milliseconds\n window_size = int(sr * frame_length * 0.001)\n padded_window_size = _next_power_of_2(\n window_size) if round_to_power_of_two else window_size\n\n assert 2 <= window_size <= len(waveform), (\n 'choose a window size {} that is [2, {}]'.format(window_size,\n len(waveform)))\n assert 0 < window_shift, '`window_shift` must be greater than 0'\n assert padded_window_size % 2 == 0, 'the padded `window_size` must be divisible by two.' \\\n ' use `round_to_power_of_two` or change `frame_length`'\n assert 0. <= preemphasis_coefficient <= 1.0, '`preemphasis_coefficient` must be between [0,1]'\n assert sr > 0, '`sr` must be greater than zero'\n return waveform, window_shift, window_size, padded_window_size\n\n\ndef _get_window(waveform: Tensor,\n padded_window_size: int,\n window_size: int,\n window_shift: int,\n window_type: str,\n blackman_coeff: float,\n snip_edges: bool,\n raw_energy: bool,\n energy_floor: float,\n dither: float,\n remove_dc_offset: bool,\n preemphasis_coefficient: float) -> Tuple[Tensor, Tensor]:\n dtype = waveform.dtype\n epsilon = _get_epsilon(dtype)\n\n # (m, window_size)\n strided_input = _get_strided(waveform, window_size, window_shift,\n snip_edges)\n\n if dither != 0.0:\n x = paddle.maximum(epsilon,\n paddle.rand(strided_input.shape, dtype=dtype))\n rand_gauss = paddle.sqrt(-2 * x.log()) * paddle.cos(2 * math.pi * x)\n strided_input = strided_input + rand_gauss * dither\n\n if remove_dc_offset:\n row_means = paddle.mean(strided_input, axis=1).unsqueeze(1) # (m, 1)\n strided_input = strided_input - row_means\n\n if raw_energy:\n signal_log_energy = _get_log_energy(strided_input, epsilon,\n energy_floor) # (m)\n\n if preemphasis_coefficient != 0.0:\n offset_strided_input = paddle.nn.functional.pad(\n strided_input.unsqueeze(0), (1, 0),\n data_format='NCL',\n mode='replicate').squeeze(0) # (m, window_size + 1)\n strided_input = strided_input - preemphasis_coefficient * offset_strided_input[:, :\n -1]\n\n window_function = _feature_window_function(\n window_type, window_size, blackman_coeff,\n dtype).unsqueeze(0) # (1, window_size)\n strided_input = strided_input * window_function # (m, window_size)\n\n # (m, padded_window_size)\n if padded_window_size != window_size:\n padding_right = padded_window_size - window_size\n strided_input = paddle.nn.functional.pad(\n strided_input.unsqueeze(0), (0, padding_right),\n data_format='NCL',\n mode='constant',\n value=0).squeeze(0)\n\n if not raw_energy:\n signal_log_energy = _get_log_energy(strided_input, epsilon,\n energy_floor) # size (m)\n\n return strided_input, signal_log_energy\n\n\ndef _subtract_column_mean(tensor: Tensor, subtract_mean: bool) -> Tensor:\n if subtract_mean:\n col_means = paddle.mean(tensor, axis=0).unsqueeze(0)\n tensor = tensor - col_means\n return tensor\n\n\ndef spectrogram(waveform: Tensor,\n blackman_coeff: float=0.42,\n channel: int=-1,\n dither: float=0.0,\n energy_floor: float=1.0,\n frame_length: float=25.0,\n frame_shift: float=10.0,\n preemphasis_coefficient: float=0.97,\n raw_energy: bool=True,\n remove_dc_offset: bool=True,\n round_to_power_of_two: bool=True,\n sr: int=16000,\n snip_edges: bool=True,\n subtract_mean: bool=False,\n window_type: str=\"povey\") -> Tensor:\n \"\"\"Compute and return a spectrogram from a waveform. The output is identical to Kaldi's.\n\n Args:\n waveform (Tensor): A waveform tensor with shape `(C, T)`.\n blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.\n channel (int, optional): Select the channel of waveform. Defaults to -1.\n dither (float, optional): Dithering constant . Defaults to 0.0.\n energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.\n frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.\n frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.\n preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.\n raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.\n remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.\n round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input\n to FFT. Defaults to True.\n sr (int, optional): Sample rate of input waveform. Defaults to 16000.\n snip_edges (bool, optional): Drop samples in the end of waveform that can't fit a signal frame when it\n is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.\n subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.\n window_type (str, optional): Choose type of window for FFT computation. Defaults to \"povey\".\n\n Returns:\n Tensor: A spectrogram tensor with shape `(m, padded_window_size // 2 + 1)` where m is the number of frames\n depends on frame_length and frame_shift.\n \"\"\"\n dtype = waveform.dtype\n epsilon = _get_epsilon(dtype)\n\n waveform, window_shift, window_size, padded_window_size = _get_waveform_and_window_properties(\n waveform, channel, sr, frame_shift, frame_length, round_to_power_of_two,\n preemphasis_coefficient)\n\n strided_input, signal_log_energy = _get_window(\n waveform, padded_window_size, window_size, window_shift, window_type,\n blackman_coeff, snip_edges, raw_energy, energy_floor, dither,\n remove_dc_offset, preemphasis_coefficient)\n\n # (m, padded_window_size // 2 + 1, 2)\n fft = paddle.fft.rfft(strided_input)\n\n power_spectrum = paddle.maximum(\n fft.abs().pow(2.), epsilon).log() # (m, padded_window_size // 2 + 1)\n power_spectrum[:, 0] = signal_log_energy\n\n power_spectrum = _subtract_column_mean(power_spectrum, subtract_mean)\n return power_spectrum\n\n\ndef _inverse_mel_scale_scalar(mel_freq: float) -> float:\n return 700.0 * (math.exp(mel_freq / 1127.0) - 1.0)\n\n\ndef _inverse_mel_scale(mel_freq: Tensor) -> Tensor:\n return 700.0 * ((mel_freq / 1127.0).exp() - 1.0)\n\n\ndef _mel_scale_scalar(freq: float) -> float:\n return 1127.0 * math.log(1.0 + freq / 700.0)\n\n\ndef _mel_scale(freq: Tensor) -> Tensor:\n return 1127.0 * (1.0 + freq / 700.0).log()\n\n\ndef _vtln_warp_freq(vtln_low_cutoff: float,\n vtln_high_cutoff: float,\n low_freq: float,\n high_freq: float,\n vtln_warp_factor: float,\n freq: Tensor) -> Tensor:\n assert vtln_low_cutoff > low_freq, 'be sure to set the vtln_low option higher than low_freq'\n assert vtln_high_cutoff < high_freq, 'be sure to set the vtln_high option lower than high_freq [or negative]'\n l = vtln_low_cutoff * max(1.0, vtln_warp_factor)\n h = vtln_high_cutoff * min(1.0, vtln_warp_factor)\n scale = 1.0 / vtln_warp_factor\n Fl = scale * l\n Fh = scale * h\n assert l > low_freq and h < high_freq\n scale_left = (Fl - low_freq) / (l - low_freq)\n scale_right = (high_freq - Fh) / (high_freq - h)\n res = paddle.empty_like(freq)\n\n outside_low_high_freq = paddle.less_than(freq, paddle.to_tensor(low_freq)) \\\n | paddle.greater_than(freq, paddle.to_tensor(high_freq))\n before_l = paddle.less_than(freq, paddle.to_tensor(l))\n before_h = paddle.less_than(freq, paddle.to_tensor(h))\n after_h = paddle.greater_equal(freq, paddle.to_tensor(h))\n\n res[after_h] = high_freq + scale_right * (freq[after_h] - high_freq)\n res[before_h] = scale * freq[before_h]\n res[before_l] = low_freq + scale_left * (freq[before_l] - low_freq)\n res[outside_low_high_freq] = freq[outside_low_high_freq]\n\n return res\n\n\ndef _vtln_warp_mel_freq(vtln_low_cutoff: float,\n vtln_high_cutoff: float,\n low_freq,\n high_freq: float,\n vtln_warp_factor: float,\n mel_freq: Tensor) -> Tensor:\n return _mel_scale(\n _vtln_warp_freq(vtln_low_cutoff, vtln_high_cutoff, low_freq, high_freq,\n vtln_warp_factor, _inverse_mel_scale(mel_freq)))\n\n\ndef _get_mel_banks(num_bins: int,\n window_length_padded: int,\n sample_freq: float,\n low_freq: float,\n high_freq: float,\n vtln_low: float,\n vtln_high: float,\n vtln_warp_factor: float) -> Tuple[Tensor, Tensor]:\n assert num_bins > 3, 'Must have at least 3 mel bins'\n assert window_length_padded % 2 == 0\n num_fft_bins = window_length_padded / 2\n nyquist = 0.5 * sample_freq\n\n if high_freq <= 0.0:\n high_freq += nyquist\n\n assert (0.0 <= low_freq < nyquist) and (0.0 < high_freq <= nyquist) and (low_freq < high_freq), \\\n ('Bad values in options: low-freq {} and high-freq {} vs. nyquist {}'.format(low_freq, high_freq, nyquist))\n\n fft_bin_width = sample_freq / window_length_padded\n mel_low_freq = _mel_scale_scalar(low_freq)\n mel_high_freq = _mel_scale_scalar(high_freq)\n\n mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1)\n\n if vtln_high < 0.0:\n vtln_high += nyquist\n\n assert vtln_warp_factor == 1.0 or ((low_freq < vtln_low < high_freq) and\n (0.0 < vtln_high < high_freq) and (vtln_low < vtln_high)), \\\n ('Bad values in options: vtln-low {} and vtln-high {}, versus '\n 'low-freq {} and high-freq {}'.format(vtln_low, vtln_high, low_freq, high_freq))\n\n bin = paddle.arange(num_bins, dtype=paddle.float32).unsqueeze(1)\n # left_mel = mel_low_freq + bin * mel_freq_delta # (num_bins, 1)\n # center_mel = mel_low_freq + (bin + 1.0) * mel_freq_delta # (num_bins, 1)\n # right_mel = mel_low_freq + (bin + 2.0) * mel_freq_delta # (num_bins, 1)\n left_mel = mel_low_freq + bin * mel_freq_delta # (num_bins, 1)\n center_mel = left_mel + mel_freq_delta\n right_mel = center_mel + mel_freq_delta\n\n if vtln_warp_factor != 1.0:\n left_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq, high_freq,\n vtln_warp_factor, left_mel)\n center_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq,\n high_freq, vtln_warp_factor,\n center_mel)\n right_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq,\n high_freq, vtln_warp_factor, right_mel)\n\n center_freqs = _inverse_mel_scale(center_mel) # (num_bins)\n # (1, num_fft_bins)\n mel = _mel_scale(fft_bin_width * paddle.arange(\n num_fft_bins, dtype=paddle.float32)).unsqueeze(0)\n\n # (num_bins, num_fft_bins)\n up_slope = (mel - left_mel) / (center_mel - left_mel)\n down_slope = (right_mel - mel) / (right_mel - center_mel)\n\n if vtln_warp_factor == 1.0:\n bins = paddle.maximum(\n paddle.zeros([1]), paddle.minimum(up_slope, down_slope))\n else:\n bins = paddle.zeros_like(up_slope)\n up_idx = paddle.greater_than(mel, left_mel) & paddle.less_than(\n mel, center_mel)\n down_idx = paddle.greater_than(mel, center_mel) & paddle.less_than(\n mel, right_mel)\n bins[up_idx] = up_slope[up_idx]\n bins[down_idx] = down_slope[down_idx]\n\n return bins, center_freqs\n\n\ndef fbank(waveform: Tensor,\n blackman_coeff: float=0.42,\n channel: int=-1,\n dither: float=0.0,\n energy_floor: float=1.0,\n frame_length: float=25.0,\n frame_shift: float=10.0,\n high_freq: float=0.0,\n htk_compat: bool=False,\n low_freq: float=20.0,\n n_mels: int=23,\n preemphasis_coefficient: float=0.97,\n raw_energy: bool=True,\n remove_dc_offset: bool=True,\n round_to_power_of_two: bool=True,\n sr: int=16000,\n snip_edges: bool=True,\n subtract_mean: bool=False,\n use_energy: bool=False,\n use_log_fbank: bool=True,\n use_power: bool=True,\n vtln_high: float=-500.0,\n vtln_low: float=100.0,\n vtln_warp: float=1.0,\n window_type: str=\"povey\") -> Tensor:\n \"\"\"Compute and return filter banks from a waveform. The output is identical to Kaldi's.\n\n Args:\n waveform (Tensor): A waveform tensor with shape `(C, T)`. `C` is in the range [0,1]. \n blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.\n channel (int, optional): Select the channel of waveform. Defaults to -1.\n dither (float, optional): Dithering constant . Defaults to 0.0.\n energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.\n frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.\n frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.\n high_freq (float, optional): The upper cut-off frequency. Defaults to 0.0.\n htk_compat (bool, optional): Put energy to the last when it is set True. Defaults to False.\n low_freq (float, optional): The lower cut-off frequency. Defaults to 20.0.\n n_mels (int, optional): Number of output mel bins. Defaults to 23.\n preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.\n raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.\n remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.\n round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input\n to FFT. Defaults to True.\n sr (int, optional): Sample rate of input waveform. Defaults to 16000.\n snip_edges (bool, optional): Drop samples in the end of waveform that can't fit a signal frame when it\n is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.\n subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.\n use_energy (bool, optional): Add an dimension with energy of spectrogram to the output. Defaults to False.\n use_log_fbank (bool, optional): Return log fbank when it is set True. Defaults to True.\n use_power (bool, optional): Whether to use power instead of magnitude. Defaults to True.\n vtln_high (float, optional): High inflection point in piecewise linear VTLN warping function. Defaults to -500.0.\n vtln_low (float, optional): Low inflection point in piecewise linear VTLN warping function. Defaults to 100.0.\n vtln_warp (float, optional): Vtln warp factor. Defaults to 1.0.\n window_type (str, optional): Choose type of window for FFT computation. Defaults to \"povey\".\n\n Returns:\n Tensor: A filter banks tensor with shape `(m, n_mels)`.\n \"\"\"\n dtype = waveform.dtype\n\n waveform, window_shift, window_size, padded_window_size = _get_waveform_and_window_properties(\n waveform, channel, sr, frame_shift, frame_length, round_to_power_of_two,\n preemphasis_coefficient)\n\n strided_input, signal_log_energy = _get_window(\n waveform, padded_window_size, window_size, window_shift, window_type,\n blackman_coeff, snip_edges, raw_energy, energy_floor, dither,\n remove_dc_offset, preemphasis_coefficient)\n\n # (m, padded_window_size // 2 + 1)\n spectrum = paddle.fft.rfft(strided_input).abs()\n if use_power:\n spectrum = spectrum.pow(2.)\n\n # (n_mels, padded_window_size // 2)\n mel_energies, _ = _get_mel_banks(n_mels, padded_window_size, sr, low_freq,\n high_freq, vtln_low, vtln_high, vtln_warp)\n # mel_energies = mel_energies.astype(dtype)\n assert mel_energies.dtype == dtype\n\n # (n_mels, padded_window_size // 2 + 1)\n mel_energies = paddle.nn.functional.pad(\n mel_energies.unsqueeze(0), (0, 1),\n data_format='NCL',\n mode='constant',\n value=0).squeeze(0)\n\n # (m, n_mels)\n mel_energies = paddle.mm(spectrum, mel_energies.T)\n if use_log_fbank:\n mel_energies = paddle.maximum(mel_energies, _get_epsilon(dtype)).log()\n\n if use_energy:\n signal_log_energy = signal_log_energy.unsqueeze(1)\n if htk_compat:\n mel_energies = paddle.concat(\n (mel_energies, signal_log_energy), axis=1)\n else:\n mel_energies = paddle.concat(\n (signal_log_energy, mel_energies), axis=1)\n\n # (m, n_mels + 1)\n mel_energies = _subtract_column_mean(mel_energies, subtract_mean)\n return mel_energies\n\n\ndef _get_dct_matrix(n_mfcc: int, n_mels: int) -> Tensor:\n dct_matrix = create_dct(n_mels, n_mels, 'ortho')\n dct_matrix[:, 0] = math.sqrt(1 / float(n_mels))\n dct_matrix = dct_matrix[:, :n_mfcc] # (n_mels, n_mfcc)\n return dct_matrix\n\n\ndef _get_lifter_coeffs(n_mfcc: int, cepstral_lifter: float) -> Tensor:\n i = paddle.arange(n_mfcc)\n return 1.0 + 0.5 * cepstral_lifter * paddle.sin(math.pi * i /\n cepstral_lifter)\n\n\ndef mfcc(waveform: Tensor,\n blackman_coeff: float=0.42,\n cepstral_lifter: float=22.0,\n channel: int=-1,\n dither: float=0.0,\n energy_floor: float=1.0,\n frame_length: float=25.0,\n frame_shift: float=10.0,\n high_freq: float=0.0,\n htk_compat: bool=False,\n low_freq: float=20.0,\n n_mfcc: int=13,\n n_mels: int=23,\n preemphasis_coefficient: float=0.97,\n raw_energy: bool=True,\n remove_dc_offset: bool=True,\n round_to_power_of_two: bool=True,\n sr: int=16000,\n snip_edges: bool=True,\n subtract_mean: bool=False,\n use_energy: bool=False,\n vtln_high: float=-500.0,\n vtln_low: float=100.0,\n vtln_warp: float=1.0,\n window_type: str=\"povey\") -> Tensor:\n \"\"\"Compute and return mel frequency cepstral coefficients from a waveform. The output is\n identical to Kaldi's.\n\n Args:\n waveform (Tensor): A waveform tensor with shape `(C, T)`.\n blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.\n cepstral_lifter (float, optional): Scaling of output mfccs. Defaults to 22.0.\n channel (int, optional): Select the channel of waveform. Defaults to -1.\n dither (float, optional): Dithering constant . Defaults to 0.0.\n energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.\n frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.\n frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.\n high_freq (float, optional): The upper cut-off frequency. Defaults to 0.0.\n htk_compat (bool, optional): Put energy to the last when it is set True. Defaults to False.\n low_freq (float, optional): The lower cut-off frequency. Defaults to 20.0.\n n_mfcc (int, optional): Number of cepstra in MFCC. Defaults to 13.\n n_mels (int, optional): Number of output mel bins. Defaults to 23.\n preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.\n raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.\n remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.\n round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input\n to FFT. Defaults to True.\n sr (int, optional): Sample rate of input waveform. Defaults to 16000.\n snip_edges (bool, optional): Drop samples in the end of waveform that can't fit a signal frame when it\n is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.\n subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.\n use_energy (bool, optional): Add an dimension with energy of spectrogram to the output. Defaults to False.\n vtln_high (float, optional): High inflection point in piecewise linear VTLN warping function. Defaults to -500.0.\n vtln_low (float, optional): Low inflection point in piecewise linear VTLN warping function. Defaults to 100.0.\n vtln_warp (float, optional): Vtln warp factor. Defaults to 1.0.\n window_type (str, optional): Choose type of window for FFT computation. Defaults to POVEY.\n\n Returns:\n Tensor: A mel frequency cepstral coefficients tensor with shape `(m, n_mfcc)`.\n \"\"\"\n assert n_mfcc <= n_mels, 'n_mfcc cannot be larger than n_mels: %d vs %d' % (\n n_mfcc, n_mels)\n\n dtype = waveform.dtype\n\n # (m, n_mels + use_energy)\n feature = fbank(\n waveform=waveform,\n blackman_coeff=blackman_coeff,\n channel=channel,\n dither=dither,\n energy_floor=energy_floor,\n frame_length=frame_length,\n frame_shift=frame_shift,\n high_freq=high_freq,\n htk_compat=htk_compat,\n low_freq=low_freq,\n n_mels=n_mels,\n preemphasis_coefficient=preemphasis_coefficient,\n raw_energy=raw_energy,\n remove_dc_offset=remove_dc_offset,\n round_to_power_of_two=round_to_power_of_two,\n sr=sr,\n snip_edges=snip_edges,\n subtract_mean=False,\n use_energy=use_energy,\n use_log_fbank=True,\n use_power=True,\n vtln_high=vtln_high,\n vtln_low=vtln_low,\n vtln_warp=vtln_warp,\n window_type=window_type)\n\n if use_energy:\n # (m)\n signal_log_energy = feature[:, n_mels if htk_compat else 0]\n mel_offset = int(not htk_compat)\n feature = feature[:, mel_offset:(n_mels + mel_offset)]\n\n # (n_mels, n_mfcc)\n dct_matrix = _get_dct_matrix(n_mfcc, n_mels).astype(dtype=dtype)\n\n # (m, n_mfcc)\n feature = feature.matmul(dct_matrix)\n\n if cepstral_lifter != 0.0:\n # (1, n_mfcc)\n lifter_coeffs = _get_lifter_coeffs(n_mfcc, cepstral_lifter).unsqueeze(0)\n feature *= lifter_coeffs.astype(dtype=dtype)\n\n if use_energy:\n feature[:, 0] = signal_log_energy\n\n if htk_compat:\n energy = feature[:, 0].unsqueeze(1) # (m, 1)\n feature = feature[:, 1:] # (m, n_mfcc - 1)\n if not use_energy:\n energy *= math.sqrt(2)\n\n feature = paddle.concat((feature, energy), axis=1)\n\n feature = _subtract_column_mean(feature, subtract_mean)\n return feature\n"
},
{
"path": "audio/paddleaudio/compliance/librosa.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n# Modified from librosa(https://github.com/librosa/librosa)\nimport warnings\nfrom typing import List\nfrom typing import Optional\nfrom typing import Union\n\nimport numpy as np\nimport scipy\nfrom numpy.lib.stride_tricks import as_strided\nfrom scipy import signal\n\nfrom ..backends import depth_convert\nfrom ..utils import ParameterError\n\n__all__ = [\n # dsp\n 'stft',\n 'mfcc',\n 'hz_to_mel',\n 'mel_to_hz',\n 'mel_frequencies',\n 'power_to_db',\n 'compute_fbank_matrix',\n 'melspectrogram',\n 'spectrogram',\n 'mu_encode',\n 'mu_decode',\n # augmentation\n 'depth_augment',\n 'spect_augment',\n 'random_crop1d',\n 'random_crop2d',\n 'adaptive_spect_augment',\n]\n\n\ndef _pad_center(data: np.ndarray, size: int, axis: int=-1,\n **kwargs) -> np.ndarray:\n \"\"\"Pad an array to a target length along a target axis.\n\n This differs from `np.pad` by centering the data prior to padding,\n analogous to `str.center`\n \"\"\"\n\n kwargs.setdefault(\"mode\", \"constant\")\n n = data.shape[axis]\n lpad = int((size - n) // 2)\n lengths = [(0, 0)] * data.ndim\n lengths[axis] = (lpad, int(size - n - lpad))\n\n if lpad < 0:\n raise ParameterError((\"Target size ({size:d}) must be \"\n \"at least input size ({n:d})\"))\n\n return np.pad(data, lengths, **kwargs)\n\n\ndef _split_frames(x: np.ndarray,\n frame_length: int,\n hop_length: int,\n axis: int=-1) -> np.ndarray:\n \"\"\"Slice a data array into (overlapping) frames.\n\n This function is aligned with librosa.frame\n \"\"\"\n\n if not isinstance(x, np.ndarray):\n raise ParameterError(\n f\"Input must be of type numpy.ndarray, given type(x)={type(x)}\")\n\n if x.shape[axis] < frame_length:\n raise ParameterError(f\"Input is too short (n={x.shape[axis]:d})\"\n f\" for frame_length={frame_length:d}\")\n\n if hop_length < 1:\n raise ParameterError(f\"Invalid hop_length: {hop_length:d}\")\n\n if axis == -1 and not x.flags[\"F_CONTIGUOUS\"]:\n warnings.warn(f\"librosa.util.frame called with axis={axis} \"\n \"on a non-contiguous input. This will result in a copy.\")\n x = np.asfortranarray(x)\n elif axis == 0 and not x.flags[\"C_CONTIGUOUS\"]:\n warnings.warn(f\"librosa.util.frame called with axis={axis} \"\n \"on a non-contiguous input. This will result in a copy.\")\n x = np.ascontiguousarray(x)\n\n n_frames = 1 + (x.shape[axis] - frame_length) // hop_length\n strides = np.asarray(x.strides)\n\n new_stride = np.prod(strides[strides > 0] // x.itemsize) * x.itemsize\n\n if axis == -1:\n shape = list(x.shape)[:-1] + [frame_length, n_frames]\n strides = list(strides) + [hop_length * new_stride]\n\n elif axis == 0:\n shape = [n_frames, frame_length] + list(x.shape)[1:]\n strides = [hop_length * new_stride] + list(strides)\n\n else:\n raise ParameterError(f\"Frame axis={axis} must be either 0 or -1\")\n\n return as_strided(x, shape=shape, strides=strides)\n\n\ndef _check_audio(y, mono=True) -> bool:\n \"\"\"Determine whether a variable contains valid audio data.\n\n The audio y must be a np.ndarray, ether 1-channel or two channel\n \"\"\"\n if not isinstance(y, np.ndarray):\n raise ParameterError(\"Audio data must be of type numpy.ndarray\")\n if y.ndim > 2:\n raise ParameterError(\n f\"Invalid shape for audio ndim={y.ndim:d}, shape={y.shape}\")\n\n if mono and y.ndim == 2:\n raise ParameterError(\n f\"Invalid shape for mono audio ndim={y.ndim:d}, shape={y.shape}\")\n\n if (mono and len(y) == 0) or (not mono and y.shape[1] < 0):\n raise ParameterError(f\"Audio is empty ndim={y.ndim:d}, shape={y.shape}\")\n\n if not np.issubdtype(y.dtype, np.floating):\n raise ParameterError(\"Audio data must be floating-point\")\n\n if not np.isfinite(y).all():\n raise ParameterError(\"Audio buffer is not finite everywhere\")\n\n return True\n\n\ndef hz_to_mel(frequencies: Union[float, List[float], np.ndarray],\n htk: bool=False) -> np.ndarray:\n \"\"\"Convert Hz to Mels.\n\n Args:\n frequencies (Union[float, List[float], np.ndarray]): Frequencies in Hz.\n htk (bool, optional): Use htk scaling. Defaults to False.\n\n Returns:\n np.ndarray: Frequency in mels.\n \"\"\"\n freq = np.asanyarray(frequencies)\n\n if htk:\n return 2595.0 * np.log10(1.0 + freq / 700.0)\n\n # Fill in the linear part\n f_min = 0.0\n f_sp = 200.0 / 3\n\n mels = (freq - f_min) / f_sp\n\n # Fill in the log-scale part\n\n min_log_hz = 1000.0 # beginning of log region (Hz)\n min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)\n logstep = np.log(6.4) / 27.0 # step size for log region\n\n if freq.ndim:\n # If we have array data, vectorize\n log_t = freq >= min_log_hz\n mels[log_t] = min_log_mel + \\\n np.log(freq[log_t] / min_log_hz) / logstep\n elif freq >= min_log_hz:\n # If we have scalar data, heck directly\n mels = min_log_mel + np.log(freq / min_log_hz) / logstep\n\n return mels\n\n\ndef mel_to_hz(mels: Union[float, List[float], np.ndarray],\n htk: int=False) -> np.ndarray:\n \"\"\"Convert mel bin numbers to frequencies.\n\n Args:\n mels (Union[float, List[float], np.ndarray]): Frequency in mels.\n htk (bool, optional): Use htk scaling. Defaults to False.\n\n Returns:\n np.ndarray: Frequencies in Hz.\n \"\"\"\n mel_array = np.asanyarray(mels)\n\n if htk:\n return 700.0 * (10.0**(mel_array / 2595.0) - 1.0)\n\n # Fill in the linear scale\n f_min = 0.0\n f_sp = 200.0 / 3\n freqs = f_min + f_sp * mel_array\n\n # And now the nonlinear scale\n min_log_hz = 1000.0 # beginning of log region (Hz)\n min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)\n logstep = np.log(6.4) / 27.0 # step size for log region\n\n if mel_array.ndim:\n # If we have vector data, vectorize\n log_t = mel_array >= min_log_mel\n freqs[log_t] = min_log_hz * \\\n np.exp(logstep * (mel_array[log_t] - min_log_mel))\n elif mel_array >= min_log_mel:\n # If we have scalar data, check directly\n freqs = min_log_hz * np.exp(logstep * (mel_array - min_log_mel))\n\n return freqs\n\n\ndef mel_frequencies(n_mels: int=128,\n fmin: float=0.0,\n fmax: float=11025.0,\n htk: bool=False) -> np.ndarray:\n \"\"\"Compute mel frequencies.\n\n Args:\n n_mels (int, optional): Number of mel bins. Defaults to 128.\n fmin (float, optional): Minimum frequency in Hz. Defaults to 0.0.\n fmax (float, optional): Maximum frequency in Hz. Defaults to 11025.0.\n htk (bool, optional): Use htk scaling. Defaults to False.\n\n Returns:\n np.ndarray: Vector of n_mels frequencies in Hz with shape `(n_mels,)`.\n \"\"\"\n # 'Center freqs' of mel bands - uniformly spaced between limits\n min_mel = hz_to_mel(fmin, htk=htk)\n max_mel = hz_to_mel(fmax, htk=htk)\n\n mels = np.linspace(min_mel, max_mel, n_mels)\n\n return mel_to_hz(mels, htk=htk)\n\n\ndef fft_frequencies(sr: int, n_fft: int) -> np.ndarray:\n \"\"\"Compute fourier frequencies.\n\n Args:\n sr (int): Sample rate.\n n_fft (int): FFT size.\n\n Returns:\n np.ndarray: FFT frequencies in Hz with shape `(n_fft//2 + 1,)`.\n \"\"\"\n return np.linspace(0, float(sr) / 2, int(1 + n_fft // 2), endpoint=True)\n\n\ndef compute_fbank_matrix(sr: int,\n n_fft: int,\n n_mels: int=128,\n fmin: float=0.0,\n fmax: Optional[float]=None,\n htk: bool=False,\n norm: str=\"slaney\",\n dtype: type=np.float32) -> np.ndarray:\n \"\"\"Compute fbank matrix.\n\n Args:\n sr (int): Sample rate.\n n_fft (int): FFT size.\n n_mels (int, optional): Number of mel bins. Defaults to 128.\n fmin (float, optional): Minimum frequency in Hz. Defaults to 0.0.\n fmax (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n htk (bool, optional): Use htk scaling. Defaults to False.\n norm (str, optional): Type of normalization. Defaults to \"slaney\".\n dtype (type, optional): Data type. Defaults to np.float32.\n\n\n Returns:\n np.ndarray: Mel transform matrix with shape `(n_mels, n_fft//2 + 1)`.\n \"\"\"\n if norm != \"slaney\":\n raise ParameterError('norm must set to slaney')\n\n if fmax is None:\n fmax = float(sr) / 2\n\n # Initialize the weights\n n_mels = int(n_mels)\n weights = np.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)\n\n # Center freqs of each FFT bin\n fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft)\n\n # 'Center freqs' of mel bands - uniformly spaced between limits\n mel_f = mel_frequencies(n_mels + 2, fmin=fmin, fmax=fmax, htk=htk)\n\n fdiff = np.diff(mel_f)\n ramps = np.subtract.outer(mel_f, fftfreqs)\n\n for i in range(n_mels):\n # lower and upper slopes for all bins\n lower = -ramps[i] / fdiff[i]\n upper = ramps[i + 2] / fdiff[i + 1]\n\n # .. then intersect them with each other and zero\n weights[i] = np.maximum(0, np.minimum(lower, upper))\n\n if norm == \"slaney\":\n # Slaney-style mel is scaled to be approx constant energy per channel\n enorm = 2.0 / (mel_f[2:n_mels + 2] - mel_f[:n_mels])\n weights *= enorm[:, np.newaxis]\n\n # Only check weights if f_mel[0] is positive\n if not np.all((mel_f[:-2] == 0) | (weights.max(axis=1) > 0)):\n # This means we have an empty channel somewhere\n warnings.warn(\"Empty filters detected in mel frequency basis. \"\n \"Some channels will produce empty responses. \"\n \"Try increasing your sampling rate (and fmax) or \"\n \"reducing n_mels.\")\n\n return weights\n\n\ndef stft(x: np.ndarray,\n n_fft: int=2048,\n hop_length: Optional[int]=None,\n win_length: Optional[int]=None,\n window: str=\"hann\",\n center: bool=True,\n dtype: type=np.complex64,\n pad_mode: str=\"reflect\") -> np.ndarray:\n \"\"\"Short-time Fourier transform (STFT).\n\n Args:\n x (np.ndarray): Input waveform in one dimension.\n n_fft (int, optional): FFT size. Defaults to 2048.\n hop_length (Optional[int], optional): Number of steps to advance between adjacent windows. Defaults to None.\n win_length (Optional[int], optional): The size of window. Defaults to None.\n window (str, optional): A string of window specification. Defaults to \"hann\".\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n dtype (type, optional): Data type of STFT results. Defaults to np.complex64.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to \"reflect\".\n\n Returns:\n np.ndarray: The complex STFT output with shape `(n_fft//2 + 1, num_frames)`.\n \"\"\"\n _check_audio(x)\n\n # By default, use the entire frame\n if win_length is None:\n win_length = n_fft\n\n # Set the default hop, if it's not already specified\n if hop_length is None:\n hop_length = int(win_length // 4)\n\n fft_window = signal.get_window(window, win_length, fftbins=True)\n\n # Pad the window out to n_fft size\n fft_window = _pad_center(fft_window, n_fft)\n\n # Reshape so that the window can be broadcast\n fft_window = fft_window.reshape((-1, 1))\n\n # Pad the time series so that frames are centered\n if center:\n if n_fft > x.shape[-1]:\n warnings.warn(\n f\"n_fft={n_fft} is too small for input signal of length={x.shape[-1]}\"\n )\n x = np.pad(x, int(n_fft // 2), mode=pad_mode)\n\n elif n_fft > x.shape[-1]:\n raise ParameterError(\n f\"n_fft={n_fft} is too small for input signal of length={x.shape[-1]}\"\n )\n\n # Window the time series.\n x_frames = _split_frames(x, frame_length=n_fft, hop_length=hop_length)\n # Pre-allocate the STFT matrix\n stft_matrix = np.empty(\n (int(1 + n_fft // 2), x_frames.shape[1]), dtype=dtype, order=\"F\")\n fft = np.fft # use numpy fft as default\n # Constrain STFT block sizes to 256 KB\n MAX_MEM_BLOCK = 2**8 * 2**10\n # how many columns can we fit within MAX_MEM_BLOCK?\n n_columns = MAX_MEM_BLOCK // (stft_matrix.shape[0] * stft_matrix.itemsize)\n n_columns = max(n_columns, 1)\n\n for bl_s in range(0, stft_matrix.shape[1], n_columns):\n bl_t = min(bl_s + n_columns, stft_matrix.shape[1])\n stft_matrix[:, bl_s:bl_t] = fft.rfft(\n fft_window * x_frames[:, bl_s:bl_t], axis=0)\n\n return stft_matrix\n\n\ndef power_to_db(spect: np.ndarray,\n ref: float=1.0,\n amin: float=1e-10,\n top_db: Optional[float]=80.0) -> np.ndarray:\n \"\"\"Convert a power spectrogram (amplitude squared) to decibel (dB) units. The function computes the scaling `10 * log10(x / ref)` in a numerically stable way.\n\n Args:\n spect (np.ndarray): STFT power spectrogram of an input waveform.\n ref (float, optional): The reference value. If smaller than 1.0, the db level of the signal will be pulled up accordingly. Otherwise, the db level is pushed down. Defaults to 1.0.\n amin (float, optional): Minimum threshold. Defaults to 1e-10.\n top_db (Optional[float], optional): Threshold the output at `top_db` below the peak. Defaults to 80.0.\n\n Returns:\n np.ndarray: Power spectrogram in db scale.\n \"\"\"\n spect = np.asarray(spect)\n\n if amin <= 0:\n raise ParameterError(\"amin must be strictly positive\")\n\n if np.issubdtype(spect.dtype, np.complexfloating):\n warnings.warn(\n \"power_to_db was called on complex input so phase \"\n \"information will be discarded. To suppress this warning, \"\n \"call power_to_db(np.abs(D)**2) instead.\")\n magnitude = np.abs(spect)\n else:\n magnitude = spect\n\n if callable(ref):\n # User supplied a function to calculate reference power\n ref_value = ref(magnitude)\n else:\n ref_value = np.abs(ref)\n\n log_spec = 10.0 * np.log10(np.maximum(amin, magnitude))\n log_spec -= 10.0 * np.log10(np.maximum(amin, ref_value))\n\n if top_db is not None:\n if top_db < 0:\n raise ParameterError(\"top_db must be non-negative\")\n log_spec = np.maximum(log_spec, log_spec.max() - top_db)\n\n return log_spec\n\n\ndef mfcc(x: np.ndarray,\n sr: int=16000,\n spect: Optional[np.ndarray]=None,\n n_mfcc: int=20,\n dct_type: int=2,\n norm: str=\"ortho\",\n lifter: int=0,\n **kwargs) -> np.ndarray:\n \"\"\"Mel-frequency cepstral coefficients (MFCCs)\n\n Args:\n x (np.ndarray): Input waveform in one dimension.\n sr (int, optional): Sample rate. Defaults to 16000.\n spect (Optional[np.ndarray], optional): Input log-power Mel spectrogram. Defaults to None.\n n_mfcc (int, optional): Number of cepstra in MFCC. Defaults to 20.\n dct_type (int, optional): Discrete cosine transform (DCT) type. Defaults to 2.\n norm (str, optional): Type of normalization. Defaults to \"ortho\".\n lifter (int, optional): Cepstral filtering. Defaults to 0.\n\n Returns:\n np.ndarray: Mel frequency cepstral coefficients array with shape `(n_mfcc, num_frames)`.\n \"\"\"\n if spect is None:\n spect = melspectrogram(x, sr=sr, **kwargs)\n\n M = scipy.fftpack.dct(spect, axis=0, type=dct_type, norm=norm)[:n_mfcc]\n\n if lifter > 0:\n factor = np.sin(np.pi * np.arange(1, 1 + n_mfcc, dtype=M.dtype) /\n lifter)\n return M * factor[:, np.newaxis]\n elif lifter == 0:\n return M\n else:\n raise ParameterError(\n f\"MFCC lifter={lifter} must be a non-negative number\")\n\n\ndef melspectrogram(x: np.ndarray,\n sr: int=16000,\n window_size: int=512,\n hop_length: int=320,\n n_mels: int=64,\n fmin: float=50.0,\n fmax: Optional[float]=None,\n window: str='hann',\n center: bool=True,\n pad_mode: str='reflect',\n power: float=2.0,\n to_db: bool=True,\n ref: float=1.0,\n amin: float=1e-10,\n top_db: Optional[float]=None) -> np.ndarray:\n \"\"\"Compute mel-spectrogram.\n\n Args:\n x (np.ndarray): Input waveform in one dimension.\n sr (int, optional): Sample rate. Defaults to 16000.\n window_size (int, optional): Size of FFT and window length. Defaults to 512.\n hop_length (int, optional): Number of steps to advance between adjacent windows. Defaults to 320.\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n fmin (float, optional): Minimum frequency in Hz. Defaults to 50.0.\n fmax (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n window (str, optional): A string of window specification. Defaults to \"hann\".\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to \"reflect\".\n power (float, optional): Exponent for the magnitude melspectrogram. Defaults to 2.0.\n to_db (bool, optional): Enable db scale. Defaults to True.\n ref (float, optional): The reference value. If smaller than 1.0, the db level of the signal will be pulled up accordingly. Otherwise, the db level is pushed down. Defaults to 1.0.\n amin (float, optional): Minimum threshold. Defaults to 1e-10.\n top_db (Optional[float], optional): Threshold the output at `top_db` below the peak. Defaults to None.\n\n Returns:\n np.ndarray: The mel-spectrogram in power scale or db scale with shape `(n_mels, num_frames)`.\n \"\"\"\n _check_audio(x, mono=True)\n if len(x) <= 0:\n raise ParameterError('The input waveform is empty')\n\n if fmax is None:\n fmax = sr // 2\n if fmin < 0 or fmin >= fmax:\n raise ParameterError('fmin and fmax must satisfy 0 np.ndarray:\n \"\"\"Compute spectrogram.\n\n Args:\n x (np.ndarray): Input waveform in one dimension.\n sr (int, optional): Sample rate. Defaults to 16000.\n window_size (int, optional): Size of FFT and window length. Defaults to 512.\n hop_length (int, optional): Number of steps to advance between adjacent windows. Defaults to 320.\n window (str, optional): A string of window specification. Defaults to \"hann\".\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to \"reflect\".\n power (float, optional): Exponent for the magnitude melspectrogram. Defaults to 2.0.\n\n Returns:\n np.ndarray: The STFT spectrogram in power scale `(n_fft//2 + 1, num_frames)`.\n \"\"\"\n\n s = stft(\n x,\n n_fft=window_size,\n hop_length=hop_length,\n win_length=window_size,\n window=window,\n center=center,\n pad_mode=pad_mode)\n\n return np.abs(s)**power\n\n\ndef mu_encode(x: np.ndarray, mu: int=255, quantized: bool=True) -> np.ndarray:\n \"\"\"Mu-law encoding. Encode waveform based on mu-law companding. When quantized is True, the result will be converted to integer in range `[0,mu-1]`. Otherwise, the resulting waveform is in range `[-1,1]`.\n\n Args:\n x (np.ndarray): The input waveform to encode.\n mu (int, optional): The endoceding parameter. Defaults to 255.\n quantized (bool, optional): If `True`, quantize the encoded values into `1 + mu` distinct integer values. Defaults to True.\n\n Returns:\n np.ndarray: The mu-law encoded waveform.\n \"\"\"\n mu = 255\n y = np.sign(x) * np.log1p(mu * np.abs(x)) / np.log1p(mu)\n if quantized:\n y = np.floor((y + 1) / 2 * mu + 0.5) # convert to [0 , mu-1]\n return y\n\n\ndef mu_decode(y: np.ndarray, mu: int=255, quantized: bool=True) -> np.ndarray:\n \"\"\"Mu-law decoding. Compute the mu-law decoding given an input code. It assumes that the input `y` is in range `[0,mu-1]` when quantize is True and `[-1,1]` otherwise.\n\n Args:\n y (np.ndarray): The encoded waveform.\n mu (int, optional): The endoceding parameter. Defaults to 255.\n quantized (bool, optional): If `True`, the input is assumed to be quantized to `1 + mu` distinct integer values. Defaults to True.\n\n Returns:\n np.ndarray: The mu-law decoded waveform.\n \"\"\"\n if mu < 1:\n raise ParameterError('mu is typically set as 2**k-1, k=1, 2, 3,...')\n\n mu = mu - 1\n if quantized: # undo the quantization\n y = y * 2 / mu - 1\n x = np.sign(y) / mu * ((1 + mu)**np.abs(y) - 1)\n return x\n\n\ndef _randint(high: int) -> int:\n \"\"\"Generate one random integer in range [0 high)\n\n This is a helper function for random data augmentation\n \"\"\"\n return int(np.random.randint(0, high=high))\n\n\ndef depth_augment(y: np.ndarray,\n choices: List=['int8', 'int16'],\n probs: List[float]=[0.5, 0.5]) -> np.ndarray:\n \"\"\" Audio depth augmentation. Do audio depth augmentation to simulate the distortion brought by quantization.\n\n Args:\n y (np.ndarray): Input waveform array in 1D or 2D.\n choices (List, optional): A list of data type to depth conversion. Defaults to ['int8', 'int16'].\n probs (List[float], optional): Probabilities to depth conversion. Defaults to [0.5, 0.5].\n\n Returns:\n np.ndarray: The augmented waveform.\n \"\"\"\n assert len(probs) == len(\n choices\n ), 'number of choices {} must be equal to size of probs {}'.format(\n len(choices), len(probs))\n depth = np.random.choice(choices, p=probs)\n src_depth = y.dtype\n y1 = depth_convert(y, depth)\n y2 = depth_convert(y1, src_depth)\n\n return y2\n\n\ndef adaptive_spect_augment(spect: np.ndarray,\n tempo_axis: int=0,\n level: float=0.1) -> np.ndarray:\n \"\"\"Do adaptive spectrogram augmentation. The level of the augmentation is govern by the parameter level, ranging from 0 to 1, with 0 represents no augmentation.\n\n Args:\n spect (np.ndarray): Input spectrogram.\n tempo_axis (int, optional): Indicate the tempo axis. Defaults to 0.\n level (float, optional): The level factor of masking. Defaults to 0.1.\n\n Returns:\n np.ndarray: The augmented spectrogram.\n \"\"\"\n assert spect.ndim == 2., 'only supports 2d tensor or numpy array'\n if tempo_axis == 0:\n nt, nf = spect.shape\n else:\n nf, nt = spect.shape\n\n time_mask_width = int(nt * level * 0.5)\n freq_mask_width = int(nf * level * 0.5)\n\n num_time_mask = int(10 * level)\n num_freq_mask = int(10 * level)\n\n if tempo_axis == 0:\n for _ in range(num_time_mask):\n start = _randint(nt - time_mask_width)\n spect[start:start + time_mask_width, :] = 0\n for _ in range(num_freq_mask):\n start = _randint(nf - freq_mask_width)\n spect[:, start:start + freq_mask_width] = 0\n else:\n for _ in range(num_time_mask):\n start = _randint(nt - time_mask_width)\n spect[:, start:start + time_mask_width] = 0\n for _ in range(num_freq_mask):\n start = _randint(nf - freq_mask_width)\n spect[start:start + freq_mask_width, :] = 0\n\n return spect\n\n\ndef spect_augment(spect: np.ndarray,\n tempo_axis: int=0,\n max_time_mask: int=3,\n max_freq_mask: int=3,\n max_time_mask_width: int=30,\n max_freq_mask_width: int=20) -> np.ndarray:\n \"\"\"Do spectrogram augmentation in both time and freq axis.\n\n Args:\n spect (np.ndarray): Input spectrogram.\n tempo_axis (int, optional): Indicate the tempo axis. Defaults to 0.\n max_time_mask (int, optional): Maximum number of time masking. Defaults to 3.\n max_freq_mask (int, optional): Maximum number of frequency masking. Defaults to 3.\n max_time_mask_width (int, optional): Maximum width of time masking. Defaults to 30.\n max_freq_mask_width (int, optional): Maximum width of frequency masking. Defaults to 20.\n\n Returns:\n np.ndarray: The augmented spectrogram.\n \"\"\"\n assert spect.ndim == 2., 'only supports 2d tensor or numpy array'\n if tempo_axis == 0:\n nt, nf = spect.shape\n else:\n nf, nt = spect.shape\n\n num_time_mask = _randint(max_time_mask)\n num_freq_mask = _randint(max_freq_mask)\n\n time_mask_width = _randint(max_time_mask_width)\n freq_mask_width = _randint(max_freq_mask_width)\n\n if tempo_axis == 0:\n for _ in range(num_time_mask):\n start = _randint(nt - time_mask_width)\n spect[start:start + time_mask_width, :] = 0\n for _ in range(num_freq_mask):\n start = _randint(nf - freq_mask_width)\n spect[:, start:start + freq_mask_width] = 0\n else:\n for _ in range(num_time_mask):\n start = _randint(nt - time_mask_width)\n spect[:, start:start + time_mask_width] = 0\n for _ in range(num_freq_mask):\n start = _randint(nf - freq_mask_width)\n spect[start:start + freq_mask_width, :] = 0\n\n return spect\n\n\ndef random_crop1d(y: np.ndarray, crop_len: int) -> np.ndarray:\n \"\"\" Random cropping on a input waveform.\n\n Args:\n y (np.ndarray): Input waveform array in 1D.\n crop_len (int): Length of waveform to crop.\n\n Returns:\n np.ndarray: The cropped waveform.\n \"\"\"\n if y.ndim != 1:\n 'only accept 1d tensor or numpy array'\n n = len(y)\n idx = _randint(n - crop_len)\n return y[idx:idx + crop_len]\n\n\ndef random_crop2d(s: np.ndarray, crop_len: int,\n tempo_axis: int=0) -> np.ndarray:\n \"\"\" Random cropping on a spectrogram.\n\n Args:\n s (np.ndarray): Input spectrogram in 2D.\n crop_len (int): Length of spectrogram to crop.\n tempo_axis (int, optional): Indicate the tempo axis. Defaults to 0.\n\n Returns:\n np.ndarray: The cropped spectrogram.\n \"\"\"\n if tempo_axis >= s.ndim:\n raise ParameterError('axis out of range')\n\n n = s.shape[tempo_axis]\n idx = _randint(high=n - crop_len)\n sli = [slice(None) for i in range(s.ndim)]\n sli[tempo_axis] = slice(idx, idx + crop_len)\n out = s[tuple(sli)]\n return out\n"
},
{
"path": "audio/paddleaudio/datasets/__init__.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\"\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom .esc50 import ESC50\nfrom .gtzan import GTZAN\nfrom .hey_snips import HeySnips\nfrom .rirs_noises import OpenRIRNoise\nfrom .tess import TESS\nfrom .urban_sound import UrbanSound8K\nfrom .voxceleb import VoxCeleb\n"
},
{
"path": "audio/paddleaudio/datasets/dataset.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom typing import List\n\nimport numpy as np\nimport paddle\n\nfrom ..backends.soundfile_backend import soundfile_load as load_audio\nfrom ..compliance.kaldi import fbank as kaldi_fbank\nfrom ..compliance.kaldi import mfcc as kaldi_mfcc\nfrom ..compliance.librosa import melspectrogram\nfrom ..compliance.librosa import mfcc\n\nfeat_funcs = {\n 'raw': None,\n 'melspectrogram': melspectrogram,\n 'mfcc': mfcc,\n 'kaldi_fbank': kaldi_fbank,\n 'kaldi_mfcc': kaldi_mfcc,\n}\n\n\nclass AudioClassificationDataset(paddle.io.Dataset):\n \"\"\"\n Base class of audio classification dataset.\n \"\"\"\n\n def __init__(self,\n files: List[str],\n labels: List[int],\n feat_type: str='raw',\n sample_rate: int=None,\n **kwargs):\n \"\"\"\n Args:\n files (:obj:`List[str]`): A list of absolute path of audio files.\n labels (:obj:`List[int]`): Labels of audio files.\n feat_type (:obj:`str`, `optional`, defaults to `raw`):\n It identifies the feature type that user wants to extract of an audio file.\n \"\"\"\n super(AudioClassificationDataset, self).__init__()\n\n if feat_type not in feat_funcs.keys():\n raise RuntimeError(\n f\"Unknown feat_type: {feat_type}, it must be one in {list(feat_funcs.keys())}\"\n )\n\n self.files = files\n self.labels = labels\n\n self.feat_type = feat_type\n self.sample_rate = sample_rate\n self.feat_config = kwargs # Pass keyword arguments to customize feature config\n\n def _get_data(self, input_file: str):\n raise NotImplementedError\n\n def _convert_to_record(self, idx):\n file, label = self.files[idx], self.labels[idx]\n\n if self.sample_rate is None:\n waveform, sample_rate = load_audio(file)\n else:\n waveform, sample_rate = load_audio(file, sr=self.sample_rate)\n\n feat_func = feat_funcs[self.feat_type]\n\n record = {}\n if self.feat_type in ['kaldi_fbank', 'kaldi_mfcc']:\n waveform = paddle.to_tensor(waveform).unsqueeze(0) # (C, T)\n record['feat'] = feat_func(\n waveform=waveform, sr=self.sample_rate, **self.feat_config)\n else:\n record['feat'] = feat_func(\n waveform, sample_rate,\n **self.feat_config) if feat_func else waveform\n record['label'] = label\n return record\n\n def __getitem__(self, idx):\n record = self._convert_to_record(idx)\n if self.feat_type in ['kaldi_fbank', 'kaldi_mfcc']:\n return self.keys[idx], record['feat'], record['label']\n else:\n return np.array(record['feat']).transpose(), np.array(\n record['label'], dtype=np.int64)\n\n def __len__(self):\n return len(self.files)\n"
},
{
"path": "audio/paddleaudio/datasets/esc50.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport os\nfrom typing import List\nfrom typing import Tuple\n\nfrom ..utils.download import download_and_decompress\nfrom ..utils.env import DATA_HOME\nfrom .dataset import AudioClassificationDataset\n\n__all__ = ['ESC50']\n\n\nclass ESC50(AudioClassificationDataset):\n \"\"\"\n The ESC-50 dataset is a labeled collection of 2000 environmental audio recordings\n suitable for benchmarking methods of environmental sound classification. The dataset\n consists of 5-second-long recordings organized into 50 semantical classes (with\n 40 examples per class)\n\n Reference:\n ESC: Dataset for Environmental Sound Classification\n http://dx.doi.org/10.1145/2733373.2806390\n \"\"\"\n\n archives = [\n {\n 'url':\n 'https://paddleaudio.bj.bcebos.com/datasets/ESC-50-master.zip',\n 'md5': '7771e4b9d86d0945acce719c7a59305a',\n },\n ]\n label_list = [\n # Animals\n 'Dog',\n 'Rooster',\n 'Pig',\n 'Cow',\n 'Frog',\n 'Cat',\n 'Hen',\n 'Insects (flying)',\n 'Sheep',\n 'Crow',\n # Natural soundscapes & water sounds\n 'Rain',\n 'Sea waves',\n 'Crackling fire',\n 'Crickets',\n 'Chirping birds',\n 'Water drops',\n 'Wind',\n 'Pouring water',\n 'Toilet flush',\n 'Thunderstorm',\n # Human, non-speech sounds\n 'Crying baby',\n 'Sneezing',\n 'Clapping',\n 'Breathing',\n 'Coughing',\n 'Footsteps',\n 'Laughing',\n 'Brushing teeth',\n 'Snoring',\n 'Drinking, sipping',\n # Interior/domestic sounds\n 'Door knock',\n 'Mouse click',\n 'Keyboard typing',\n 'Door, wood creaks',\n 'Can opening',\n 'Washing machine',\n 'Vacuum cleaner',\n 'Clock alarm',\n 'Clock tick',\n 'Glass breaking',\n # Exterior/urban noises\n 'Helicopter',\n 'Chainsaw',\n 'Siren',\n 'Car horn',\n 'Engine',\n 'Train',\n 'Church bells',\n 'Airplane',\n 'Fireworks',\n 'Hand saw',\n ]\n meta = os.path.join('ESC-50-master', 'meta', 'esc50.csv')\n meta_info = collections.namedtuple(\n 'META_INFO',\n ('filename', 'fold', 'target', 'category', 'esc10', 'src_file', 'take'))\n audio_path = os.path.join('ESC-50-master', 'audio')\n\n def __init__(self,\n mode: str='train',\n split: int=1,\n feat_type: str='raw',\n **kwargs):\n \"\"\"\n Args:\n mode (:obj:`str`, `optional`, defaults to `train`):\n It identifies the dataset mode (train or dev).\n split (:obj:`int`, `optional`, defaults to 1):\n It specify the fold of dev dataset.\n feat_type (:obj:`str`, `optional`, defaults to `raw`):\n It identifies the feature type that user wants to extract of an audio file.\n \"\"\"\n files, labels = self._get_data(mode, split)\n super(ESC50, self).__init__(\n files=files, labels=labels, feat_type=feat_type, **kwargs)\n\n def _get_meta_info(self) -> List[collections.namedtuple]:\n ret = []\n with open(os.path.join(DATA_HOME, self.meta), 'r') as rf:\n for line in rf.readlines()[1:]:\n ret.append(self.meta_info(*line.strip().split(',')))\n return ret\n\n def _get_data(self, mode: str, split: int) -> Tuple[List[str], List[int]]:\n if not os.path.isdir(os.path.join(DATA_HOME, self.audio_path)) or \\\n not os.path.isfile(os.path.join(DATA_HOME, self.meta)):\n download_and_decompress(self.archives, DATA_HOME)\n\n meta_info = self._get_meta_info()\n\n files = []\n labels = []\n for sample in meta_info:\n filename, fold, target, _, _, _, _ = sample\n if mode == 'train' and int(fold) != split:\n files.append(os.path.join(DATA_HOME, self.audio_path, filename))\n labels.append(int(target))\n\n if mode != 'train' and int(fold) == split:\n files.append(os.path.join(DATA_HOME, self.audio_path, filename))\n labels.append(int(target))\n\n return files, labels\n"
},
{
"path": "audio/paddleaudio/datasets/gtzan.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport os\nimport random\nfrom typing import List\nfrom typing import Tuple\n\nfrom ..utils.download import download_and_decompress\nfrom ..utils.env import DATA_HOME\nfrom .dataset import AudioClassificationDataset\n\n__all__ = ['GTZAN']\n\n\nclass GTZAN(AudioClassificationDataset):\n \"\"\"\n The GTZAN dataset consists of 1000 audio tracks each 30 seconds long. It contains 10 genres,\n each represented by 100 tracks. The dataset is the most-used public dataset for evaluation\n in machine listening research for music genre recognition (MGR).\n\n Reference:\n Musical genre classification of audio signals\n https://ieeexplore.ieee.org/document/1021072/\n \"\"\"\n\n archives = [\n {\n 'url': 'http://opihi.cs.uvic.ca/sound/genres.tar.gz',\n 'md5': '5b3d6dddb579ab49814ab86dba69e7c7',\n },\n ]\n label_list = [\n 'blues', 'classical', 'country', 'disco', 'hiphop', 'jazz', 'metal',\n 'pop', 'reggae', 'rock'\n ]\n meta = os.path.join('genres', 'input.mf')\n meta_info = collections.namedtuple('META_INFO', ('file_path', 'label'))\n audio_path = 'genres'\n\n def __init__(self,\n mode='train',\n seed=0,\n n_folds=5,\n split=1,\n feat_type='raw',\n **kwargs):\n \"\"\"\n Args:\n mode (:obj:`str`, `optional`, defaults to `train`):\n It identifies the dataset mode (train or dev).\n seed (:obj:`int`, `optional`, defaults to 0):\n Set the random seed to shuffle samples.\n n_folds (:obj:`int`, `optional`, defaults to 5):\n Split the dataset into n folds. 1 fold for dev dataset and n-1 for train dataset.\n split (:obj:`int`, `optional`, defaults to 1):\n It specify the fold of dev dataset.\n feat_type (:obj:`str`, `optional`, defaults to `raw`):\n It identifies the feature type that user wants to extract of an audio file.\n \"\"\"\n assert split <= n_folds, f'The selected split should not be larger than n_fold, but got {split} > {n_folds}'\n files, labels = self._get_data(mode, seed, n_folds, split)\n super(GTZAN, self).__init__(\n files=files, labels=labels, feat_type=feat_type, **kwargs)\n\n def _get_meta_info(self) -> List[collections.namedtuple]:\n ret = []\n with open(os.path.join(DATA_HOME, self.meta), 'r') as rf:\n for line in rf.readlines():\n ret.append(self.meta_info(*line.strip().split('\\t')))\n return ret\n\n def _get_data(self, mode, seed, n_folds,\n split) -> Tuple[List[str], List[int]]:\n if not os.path.isdir(os.path.join(DATA_HOME, self.audio_path)) or \\\n not os.path.isfile(os.path.join(DATA_HOME, self.meta)):\n download_and_decompress(self.archives, DATA_HOME)\n\n meta_info = self._get_meta_info()\n random.seed(seed) # shuffle samples to split data\n random.shuffle(\n meta_info\n ) # make sure using the same seed to create train and dev dataset\n\n files = []\n labels = []\n n_samples_per_fold = len(meta_info) // n_folds\n for idx, sample in enumerate(meta_info):\n file_path, label = sample\n filename = os.path.basename(file_path)\n target = self.label_list.index(label)\n fold = idx // n_samples_per_fold + 1\n\n if mode == 'train' and int(fold) != split:\n files.append(\n os.path.join(DATA_HOME, self.audio_path, label, filename))\n labels.append(target)\n\n if mode != 'train' and int(fold) == split:\n files.append(\n os.path.join(DATA_HOME, self.audio_path, label, filename))\n labels.append(target)\n\n return files, labels\n"
},
{
"path": "audio/paddleaudio/datasets/hey_snips.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport json\nimport os\nfrom typing import List\nfrom typing import Tuple\n\nfrom .dataset import AudioClassificationDataset\n\n__all__ = ['HeySnips']\n\n\nclass HeySnips(AudioClassificationDataset):\n meta_info = collections.namedtuple('META_INFO',\n ('key', 'label', 'duration', 'wav'))\n\n def __init__(self,\n data_dir: os.PathLike,\n mode: str='train',\n feat_type: str='kaldi_fbank',\n sample_rate: int=16000,\n **kwargs):\n self.data_dir = data_dir\n files, labels = self._get_data(mode)\n super(HeySnips, self).__init__(\n files=files,\n labels=labels,\n feat_type=feat_type,\n sample_rate=sample_rate,\n **kwargs)\n\n def _get_meta_info(self, mode) -> List[collections.namedtuple]:\n ret = []\n with open(os.path.join(self.data_dir, '{}.json'.format(mode)),\n 'r') as f:\n data = json.load(f)\n for item in data:\n sample = collections.OrderedDict()\n if item['duration'] > 0:\n sample['key'] = item['id']\n sample['label'] = 0 if item['is_hotword'] == 1 else -1\n sample['duration'] = item['duration']\n sample['wav'] = os.path.join(self.data_dir,\n item['audio_file_path'])\n ret.append(self.meta_info(*sample.values()))\n return ret\n\n def _get_data(self, mode: str) -> Tuple[List[str], List[int]]:\n meta_info = self._get_meta_info(mode)\n\n files = []\n labels = []\n self.keys = []\n self.durations = []\n for sample in meta_info:\n key, target, duration, wav = sample\n files.append(wav)\n labels.append(int(target))\n self.keys.append(key)\n self.durations.append(float(duration))\n\n return files, labels\n"
},
{
"path": "audio/paddleaudio/datasets/rirs_noises.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport csv\nimport os\nimport random\nfrom typing import List\n\nfrom paddle.io import Dataset\nfrom tqdm import tqdm\n\nfrom ..backends.soundfile_backend import soundfile_load as load_audio\nfrom ..backends.soundfile_backend import soundfile_save as save_wav\nfrom ..utils import DATA_HOME\nfrom ..utils.download import download_and_decompress\nfrom .dataset import feat_funcs\n\n__all__ = ['OpenRIRNoise']\n\n\nclass OpenRIRNoise(Dataset):\n archives = [\n {\n 'url': 'http://www.openslr.org/resources/28/rirs_noises.zip',\n 'md5': 'e6f48e257286e05de56413b4779d8ffb',\n },\n ]\n\n sample_rate = 16000\n meta_info = collections.namedtuple('META_INFO', ('id', 'duration', 'wav'))\n base_path = os.path.join(DATA_HOME, 'open_rir_noise')\n wav_path = os.path.join(base_path, 'RIRS_NOISES')\n csv_path = os.path.join(base_path, 'csv')\n subsets = ['rir', 'noise']\n\n def __init__(self,\n subset: str='rir',\n feat_type: str='raw',\n target_dir=None,\n random_chunk: bool=True,\n chunk_duration: float=3.0,\n seed: int=0,\n **kwargs):\n\n assert subset in self.subsets, \\\n 'Dataset subset must be one in {}, but got {}'.format(self.subsets, subset)\n\n self.subset = subset\n self.feat_type = feat_type\n self.feat_config = kwargs\n self.random_chunk = random_chunk\n self.chunk_duration = chunk_duration\n\n OpenRIRNoise.csv_path = os.path.join(\n target_dir, \"open_rir_noise\",\n \"csv\") if target_dir else self.csv_path\n self._data = self._get_data()\n super(OpenRIRNoise, self).__init__()\n\n # Set up a seed to reproduce training or predicting result.\n # random.seed(seed)\n\n def _get_data(self):\n # Download audio files.\n print(f\"rirs noises base path: {self.base_path}\")\n if not os.path.isdir(self.base_path):\n download_and_decompress(\n self.archives, self.base_path, decompress=True)\n else:\n print(\n f\"{self.base_path} already exists, we will not download and decompress again\"\n )\n\n # Data preparation.\n print(f\"prepare the csv to {self.csv_path}\")\n if not os.path.isdir(self.csv_path):\n os.makedirs(self.csv_path)\n self.prepare_data()\n\n data = []\n with open(os.path.join(self.csv_path, f'{self.subset}.csv'), 'r') as rf:\n for line in rf.readlines()[1:]:\n audio_id, duration, wav = line.strip().split(',')\n data.append(self.meta_info(audio_id, float(duration), wav))\n\n random.shuffle(data)\n return data\n\n def _convert_to_record(self, idx: int):\n sample = self._data[idx]\n\n record = {}\n # To show all fields in a namedtuple: `type(sample)._fields`\n for field in type(sample)._fields:\n record[field] = getattr(sample, field)\n\n waveform, sr = load_audio(record['wav'])\n\n assert self.feat_type in feat_funcs.keys(), \\\n f\"Unknown feat_type: {self.feat_type}, it must be one in {list(feat_funcs.keys())}\"\n feat_func = feat_funcs[self.feat_type]\n feat = feat_func(\n waveform, sr=sr, **self.feat_config) if feat_func else waveform\n\n record.update({'feat': feat})\n return record\n\n @staticmethod\n def _get_chunks(seg_dur, audio_id, audio_duration):\n num_chunks = int(audio_duration / seg_dur) # all in milliseconds\n\n chunk_lst = [\n audio_id + \"_\" + str(i * seg_dur) + \"_\" + str(i * seg_dur + seg_dur)\n for i in range(num_chunks)\n ]\n return chunk_lst\n\n def _get_audio_info(self, wav_file: str,\n split_chunks: bool) -> List[List[str]]:\n waveform, sr = load_audio(wav_file)\n audio_id = wav_file.split(\"/open_rir_noise/\")[-1].split(\".\")[0]\n audio_duration = waveform.shape[0] / sr\n\n ret = []\n if split_chunks and audio_duration > self.chunk_duration: # Split into pieces of self.chunk_duration seconds.\n uniq_chunks_list = self._get_chunks(self.chunk_duration, audio_id,\n audio_duration)\n\n for idx, chunk in enumerate(uniq_chunks_list):\n s, e = chunk.split(\"_\")[-2:] # Timestamps of start and end\n start_sample = int(float(s) * sr)\n end_sample = int(float(e) * sr)\n new_wav_file = os.path.join(self.base_path,\n audio_id + f'_chunk_{idx+1:02}.wav')\n save_wav(waveform[start_sample:end_sample], sr, new_wav_file)\n # id, duration, new_wav\n ret.append([chunk, self.chunk_duration, new_wav_file])\n else: # Keep whole audio.\n ret.append([audio_id, audio_duration, wav_file])\n return ret\n\n def generate_csv(self,\n wav_files: List[str],\n output_file: str,\n split_chunks: bool=True):\n print(f'Generating csv: {output_file}')\n header = [\"id\", \"duration\", \"wav\"]\n\n infos = list(\n tqdm(\n map(self._get_audio_info, wav_files, [split_chunks] * len(\n wav_files)),\n total=len(wav_files)))\n\n csv_lines = []\n for info in infos:\n csv_lines.extend(info)\n\n with open(output_file, mode=\"w\") as csv_f:\n csv_writer = csv.writer(\n csv_f, delimiter=\",\", quotechar='\"', quoting=csv.QUOTE_MINIMAL)\n csv_writer.writerow(header)\n for line in csv_lines:\n csv_writer.writerow(line)\n\n def prepare_data(self):\n rir_list = os.path.join(self.wav_path, \"real_rirs_isotropic_noises\",\n \"rir_list\")\n rir_files = []\n with open(rir_list, 'r') as f:\n for line in f.readlines():\n rir_file = line.strip().split(' ')[-1]\n rir_files.append(os.path.join(self.base_path, rir_file))\n\n noise_list = os.path.join(self.wav_path, \"pointsource_noises\",\n \"noise_list\")\n noise_files = []\n with open(noise_list, 'r') as f:\n for line in f.readlines():\n noise_file = line.strip().split(' ')[-1]\n noise_files.append(os.path.join(self.base_path, noise_file))\n\n self.generate_csv(rir_files, os.path.join(self.csv_path, 'rir.csv'))\n self.generate_csv(noise_files, os.path.join(self.csv_path, 'noise.csv'))\n\n def __getitem__(self, idx):\n return self._convert_to_record(idx)\n\n def __len__(self):\n return len(self._data)\n"
},
{
"path": "audio/paddleaudio/datasets/tess.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport os\nimport random\nfrom typing import List\nfrom typing import Tuple\n\nfrom ..utils.download import download_and_decompress\nfrom ..utils.env import DATA_HOME\nfrom .dataset import AudioClassificationDataset\n\n__all__ = ['TESS']\n\n\nclass TESS(AudioClassificationDataset):\n \"\"\"\n TESS is a set of 200 target words were spoken in the carrier phrase\n \"Say the word _____' by two actresses (aged 26 and 64 years) and\n recordings were made of the set portraying each of seven emotions(anger,\n disgust, fear, happiness, pleasant surprise, sadness, and neutral).\n There are 2800 stimuli in total.\n\n Reference:\n Toronto emotional speech set (TESS)\n https://doi.org/10.5683/SP2/E8H2MF\n \"\"\"\n\n archives = [\n {\n 'url':\n 'https://bj.bcebos.com/paddleaudio/datasets/TESS_Toronto_emotional_speech_set.zip',\n 'md5':\n '1465311b24d1de704c4c63e4ccc470c7',\n },\n ]\n label_list = [\n 'angry',\n 'disgust',\n 'fear',\n 'happy',\n 'neutral',\n 'ps', # pleasant surprise\n 'sad',\n ]\n meta_info = collections.namedtuple('META_INFO',\n ('speaker', 'word', 'emotion'))\n audio_path = 'TESS_Toronto_emotional_speech_set'\n\n def __init__(self,\n mode='train',\n seed=0,\n n_folds=5,\n split=1,\n feat_type='raw',\n **kwargs):\n \"\"\"\n Args:\n mode (:obj:`str`, `optional`, defaults to `train`):\n It identifies the dataset mode (train or dev).\n seed (:obj:`int`, `optional`, defaults to 0):\n Set the random seed to shuffle samples.\n n_folds (:obj:`int`, `optional`, defaults to 5):\n Split the dataset into n folds. 1 fold for dev dataset and n-1 for train dataset.\n split (:obj:`int`, `optional`, defaults to 1):\n It specify the fold of dev dataset.\n feat_type (:obj:`str`, `optional`, defaults to `raw`):\n It identifies the feature type that user wants to extract of an audio file.\n \"\"\"\n assert split <= n_folds, f'The selected split should not be larger than n_fold, but got {split} > {n_folds}'\n files, labels = self._get_data(mode, seed, n_folds, split)\n super(TESS, self).__init__(\n files=files, labels=labels, feat_type=feat_type, **kwargs)\n\n def _get_meta_info(self, files) -> List[collections.namedtuple]:\n ret = []\n for file in files:\n basename_without_extend = os.path.basename(file)[:-4]\n ret.append(self.meta_info(*basename_without_extend.split('_')))\n return ret\n\n def _get_data(self, mode, seed, n_folds,\n split) -> Tuple[List[str], List[int]]:\n if not os.path.isdir(os.path.join(DATA_HOME, self.audio_path)):\n download_and_decompress(self.archives, DATA_HOME)\n\n wav_files = []\n for root, _, files in os.walk(os.path.join(DATA_HOME, self.audio_path)):\n for file in files:\n if file.endswith('.wav'):\n wav_files.append(os.path.join(root, file))\n\n random.seed(seed) # shuffle samples to split data\n random.shuffle(\n wav_files\n ) # make sure using the same seed to create train and dev dataset\n meta_info = self._get_meta_info(wav_files)\n\n files = []\n labels = []\n n_samples_per_fold = len(meta_info) // n_folds\n for idx, sample in enumerate(meta_info):\n _, _, emotion = sample\n target = self.label_list.index(emotion)\n fold = idx // n_samples_per_fold + 1\n\n if mode == 'train' and int(fold) != split:\n files.append(wav_files[idx])\n labels.append(target)\n\n if mode != 'train' and int(fold) == split:\n files.append(wav_files[idx])\n labels.append(target)\n\n return files, labels\n"
},
{
"path": "audio/paddleaudio/datasets/urban_sound.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport os\nfrom typing import List\nfrom typing import Tuple\n\nfrom ..utils.download import download_and_decompress\nfrom ..utils.env import DATA_HOME\nfrom .dataset import AudioClassificationDataset\n\n__all__ = ['UrbanSound8K']\n\n\nclass UrbanSound8K(AudioClassificationDataset):\n \"\"\"\n UrbanSound8K dataset contains 8732 labeled sound excerpts (<=4s) of urban\n sounds from 10 classes: air_conditioner, car_horn, children_playing, dog_bark,\n drilling, enginge_idling, gun_shot, jackhammer, siren, and street_music. The\n classes are drawn from the urban sound taxonomy.\n\n Reference:\n A Dataset and Taxonomy for Urban Sound Research\n https://dl.acm.org/doi/10.1145/2647868.2655045\n \"\"\"\n\n archives = [\n {\n 'url':\n 'https://zenodo.org/record/1203745/files/UrbanSound8K.tar.gz',\n 'md5': '9aa69802bbf37fb986f71ec1483a196e',\n },\n ]\n label_list = [\n \"air_conditioner\", \"car_horn\", \"children_playing\", \"dog_bark\",\n \"drilling\", \"engine_idling\", \"gun_shot\", \"jackhammer\", \"siren\",\n \"street_music\"\n ]\n meta = os.path.join('UrbanSound8K', 'metadata', 'UrbanSound8K.csv')\n meta_info = collections.namedtuple(\n 'META_INFO', ('filename', 'fsid', 'start', 'end', 'salience', 'fold',\n 'class_id', 'label'))\n audio_path = os.path.join('UrbanSound8K', 'audio')\n\n def __init__(self,\n mode: str='train',\n split: int=1,\n feat_type: str='raw',\n **kwargs):\n files, labels = self._get_data(mode, split)\n super(UrbanSound8K, self).__init__(\n files=files, labels=labels, feat_type=feat_type, **kwargs)\n \"\"\"\n Args:\n mode (:obj:`str`, `optional`, defaults to `train`):\n It identifies the dataset mode (train or dev).\n split (:obj:`int`, `optional`, defaults to 1):\n It specify the fold of dev dataset.\n feat_type (:obj:`str`, `optional`, defaults to `raw`):\n It identifies the feature type that user wants to extract of an audio file.\n \"\"\"\n\n def _get_meta_info(self):\n ret = []\n with open(os.path.join(DATA_HOME, self.meta), 'r') as rf:\n for line in rf.readlines()[1:]:\n ret.append(self.meta_info(*line.strip().split(',')))\n return ret\n\n def _get_data(self, mode: str, split: int) -> Tuple[List[str], List[int]]:\n if not os.path.isdir(os.path.join(DATA_HOME, self.audio_path)) or \\\n not os.path.isfile(os.path.join(DATA_HOME, self.meta)):\n download_and_decompress(self.archives, DATA_HOME)\n\n meta_info = self._get_meta_info()\n\n files = []\n labels = []\n for sample in meta_info:\n filename, _, _, _, _, fold, target, _ = sample\n if mode == 'train' and int(fold) != split:\n files.append(\n os.path.join(DATA_HOME, self.audio_path, f'fold{fold}',\n filename))\n labels.append(int(target))\n\n if mode != 'train' and int(fold) == split:\n files.append(\n os.path.join(DATA_HOME, self.audio_path, f'fold{fold}',\n filename))\n labels.append(int(target))\n\n return files, labels\n"
},
{
"path": "audio/paddleaudio/datasets/voxceleb.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport collections\nimport csv\nimport glob\nimport os\nimport random\nfrom multiprocessing import cpu_count\nfrom typing import List\n\nfrom paddle.io import Dataset\nfrom pathos.multiprocessing import Pool\nfrom tqdm import tqdm\n\nfrom ..backends.soundfile_backend import soundfile_load as load_audio\nfrom ..utils import DATA_HOME\nfrom ..utils import decompress\nfrom ..utils.download import download_and_decompress\nfrom .dataset import feat_funcs\n\n__all__ = ['VoxCeleb']\n\n\nclass VoxCeleb(Dataset):\n source_url = 'https://thor.robots.ox.ac.uk/~vgg/data/voxceleb/vox1a/'\n archives_audio_dev = [\n {\n 'url': source_url + 'vox1_dev_wav_partaa',\n 'md5': 'e395d020928bc15670b570a21695ed96',\n },\n {\n 'url': source_url + 'vox1_dev_wav_partab',\n 'md5': 'bbfaaccefab65d82b21903e81a8a8020',\n },\n {\n 'url': source_url + 'vox1_dev_wav_partac',\n 'md5': '017d579a2a96a077f40042ec33e51512',\n },\n {\n 'url': source_url + 'vox1_dev_wav_partad',\n 'md5': '7bb1e9f70fddc7a678fa998ea8b3ba19',\n },\n ]\n archives_audio_test = [\n {\n 'url': source_url + 'vox1_test_wav.zip',\n 'md5': '185fdc63c3c739954633d50379a3d102',\n },\n ]\n archives_meta = [\n {\n 'url':\n 'https://www.robots.ox.ac.uk/~vgg/data/voxceleb/meta/veri_test2.txt',\n 'md5':\n 'b73110731c9223c1461fe49cb48dddfc',\n },\n ]\n\n num_speakers = 1211 # 1211 vox1, 5994 vox2, 7205 vox1+2, test speakers: 41\n sample_rate = 16000\n meta_info = collections.namedtuple(\n 'META_INFO', ('id', 'duration', 'wav', 'start', 'stop', 'spk_id'))\n base_path = os.path.join(DATA_HOME, 'vox1')\n wav_path = os.path.join(base_path, 'wav')\n meta_path = os.path.join(base_path, 'meta')\n veri_test_file = os.path.join(meta_path, 'veri_test2.txt')\n csv_path = os.path.join(base_path, 'csv')\n subsets = ['train', 'dev', 'enroll', 'test']\n\n def __init__(\n self,\n subset: str='train',\n feat_type: str='raw',\n random_chunk: bool=True,\n chunk_duration: float=3.0, # seconds\n split_ratio: float=0.9, # train split ratio\n seed: int=0,\n target_dir: str=None,\n vox2_base_path=None,\n **kwargs):\n \"\"\"VoxCeleb data prepare and get the specific dataset audio info\n\n Args:\n subset (str, optional): dataset name, such as train, dev, enroll or test. Defaults to 'train'.\n feat_type (str, optional): feat type, such raw, melspectrogram(fbank) or mfcc . Defaults to 'raw'.\n random_chunk (bool, optional): random select a duration from audio. Defaults to True.\n chunk_duration (float, optional): chunk duration if random_chunk flag is set. Defaults to 3.0.\n target_dir (str, optional): data dir, audio info will be stored in this directory. Defaults to None.\n vox2_base_path (_type_, optional): vox2 directory. vox2 data must be converted from m4a to wav. Defaults to None.\n \"\"\"\n assert subset in self.subsets, \\\n 'Dataset subset must be one in {}, but got {}'.format(self.subsets, subset)\n\n self.subset = subset\n self.spk_id2label = {}\n self.feat_type = feat_type\n self.feat_config = kwargs\n self.random_chunk = random_chunk\n self.chunk_duration = chunk_duration\n self.split_ratio = split_ratio\n self.target_dir = target_dir if target_dir else VoxCeleb.base_path\n self.vox2_base_path = vox2_base_path\n\n # if we set the target dir, we will change the vox data info data from base path to target dir\n VoxCeleb.csv_path = os.path.join(\n target_dir, \"voxceleb\", 'csv') if target_dir else VoxCeleb.csv_path\n VoxCeleb.meta_path = os.path.join(\n target_dir, \"voxceleb\",\n 'meta') if target_dir else VoxCeleb.meta_path\n VoxCeleb.veri_test_file = os.path.join(VoxCeleb.meta_path,\n 'veri_test2.txt')\n # self._data = self._get_data()[:1000] # KP: Small dataset test.\n self._data = self._get_data()\n super(VoxCeleb, self).__init__()\n\n # Set up a seed to reproduce training or predicting result.\n # random.seed(seed)\n\n def _get_data(self):\n # Download audio files.\n # We need the users to decompress all vox1/dev/wav and vox1/test/wav/ to vox1/wav/ dir\n # so, we check the vox1/wav dir status\n print(f\"wav base path: {self.wav_path}\")\n if not os.path.isdir(self.wav_path):\n print(\"start to download the voxceleb1 dataset\")\n download_and_decompress( # multi-zip parts concatenate to vox1_dev_wav.zip\n self.archives_audio_dev,\n self.base_path,\n decompress=False)\n download_and_decompress( # download the vox1_test_wav.zip and unzip\n self.archives_audio_test,\n self.base_path,\n decompress=True)\n\n # Download all parts and concatenate the files into one zip file.\n dev_zipfile = os.path.join(self.base_path, 'vox1_dev_wav.zip')\n print(f'Concatenating all parts to: {dev_zipfile}')\n os.system(\n f'cat {os.path.join(self.base_path, \"vox1_dev_wav_parta*\")} > {dev_zipfile}'\n )\n\n # Extract all audio files of dev and test set.\n decompress(dev_zipfile, self.base_path)\n\n # Download meta files.\n if not os.path.isdir(self.meta_path):\n print(\"prepare the meta data\")\n download_and_decompress(\n self.archives_meta, self.meta_path, decompress=False)\n\n # Data preparation.\n if not os.path.isdir(self.csv_path):\n os.makedirs(self.csv_path)\n self.prepare_data()\n\n data = []\n print(\n f\"read the {self.subset} from {os.path.join(self.csv_path, f'{self.subset}.csv')}\"\n )\n with open(os.path.join(self.csv_path, f'{self.subset}.csv'), 'r') as rf:\n for line in rf.readlines()[1:]:\n audio_id, duration, wav, start, stop, spk_id = line.strip(\n ).split(',')\n data.append(\n self.meta_info(audio_id,\n float(duration), wav,\n int(start), int(stop), spk_id))\n\n with open(os.path.join(self.meta_path, 'spk_id2label.txt'), 'r') as f:\n for line in f.readlines():\n spk_id, label = line.strip().split(' ')\n self.spk_id2label[spk_id] = int(label)\n\n return data\n\n def _convert_to_record(self, idx: int):\n sample = self._data[idx]\n\n record = {}\n # To show all fields in a namedtuple: `type(sample)._fields`\n for field in type(sample)._fields:\n record[field] = getattr(sample, field)\n\n waveform, sr = load_audio(record['wav'])\n\n # random select a chunk audio samples from the audio\n if self.random_chunk:\n num_wav_samples = waveform.shape[0]\n num_chunk_samples = int(self.chunk_duration * sr)\n start = random.randint(0, num_wav_samples - num_chunk_samples - 1)\n stop = start + num_chunk_samples\n else:\n start = record['start']\n stop = record['stop']\n\n waveform = waveform[start:stop]\n\n assert self.feat_type in feat_funcs.keys(), \\\n f\"Unknown feat_type: {self.feat_type}, it must be one in {list(feat_funcs.keys())}\"\n feat_func = feat_funcs[self.feat_type]\n feat = feat_func(\n waveform, sr=sr, **self.feat_config) if feat_func else waveform\n\n record.update({'feat': feat})\n if self.subset in ['train',\n 'dev']: # Labels are available in train and dev.\n record.update({'label': self.spk_id2label[record['spk_id']]})\n\n return record\n\n @staticmethod\n def _get_chunks(seg_dur, audio_id, audio_duration):\n num_chunks = int(audio_duration / seg_dur) # all in milliseconds\n\n chunk_lst = [\n audio_id + \"_\" + str(i * seg_dur) + \"_\" + str(i * seg_dur + seg_dur)\n for i in range(num_chunks)\n ]\n return chunk_lst\n\n def _get_audio_info(self, wav_file: str,\n split_chunks: bool) -> List[List[str]]:\n waveform, sr = load_audio(wav_file)\n spk_id, sess_id, utt_id = wav_file.split(\"/\")[-3:]\n audio_id = '-'.join([spk_id, sess_id, utt_id.split(\".\")[0]])\n audio_duration = waveform.shape[0] / sr\n\n ret = []\n if split_chunks: # Split into pieces of self.chunk_duration seconds.\n uniq_chunks_list = self._get_chunks(self.chunk_duration, audio_id,\n audio_duration)\n\n for chunk in uniq_chunks_list:\n s, e = chunk.split(\"_\")[-2:] # Timestamps of start and end\n start_sample = int(float(s) * sr)\n end_sample = int(float(e) * sr)\n # id, duration, wav, start, stop, spk_id\n ret.append([\n chunk, audio_duration, wav_file, start_sample, end_sample,\n spk_id\n ])\n else: # Keep whole audio.\n ret.append([\n audio_id, audio_duration, wav_file, 0, waveform.shape[0], spk_id\n ])\n return ret\n\n def generate_csv(self,\n wav_files: List[str],\n output_file: str,\n split_chunks: bool=True):\n print(f'Generating csv: {output_file}')\n header = [\"id\", \"duration\", \"wav\", \"start\", \"stop\", \"spk_id\"]\n # Note: this may occurs c++ exception, but the program will execute fine\n # so we can ignore the exception \n with Pool(cpu_count()) as p:\n infos = list(\n tqdm(\n p.imap(lambda x: self._get_audio_info(x, split_chunks),\n wav_files),\n total=len(wav_files)))\n\n csv_lines = []\n for info in infos:\n csv_lines.extend(info)\n\n with open(output_file, mode=\"w\") as csv_f:\n csv_writer = csv.writer(\n csv_f, delimiter=\",\", quotechar='\"', quoting=csv.QUOTE_MINIMAL)\n csv_writer.writerow(header)\n for line in csv_lines:\n csv_writer.writerow(line)\n\n def prepare_data(self):\n # Audio of speakers in veri_test_file should not be included in training set.\n print(\"start to prepare the data csv file\")\n enroll_files = set()\n test_files = set()\n # get the enroll and test audio file path\n with open(self.veri_test_file, 'r') as f:\n for line in f.readlines():\n _, enrol_file, test_file = line.strip().split(' ')\n enroll_files.add(os.path.join(self.wav_path, enrol_file))\n test_files.add(os.path.join(self.wav_path, test_file))\n enroll_files = sorted(enroll_files)\n test_files = sorted(test_files)\n\n # get the enroll and test speakers\n test_spks = set()\n for file in (enroll_files + test_files):\n spk = file.split('/wav/')[1].split('/')[0]\n test_spks.add(spk)\n\n # get all the train and dev audios file path\n audio_files = []\n speakers = set()\n print(\"Getting file list...\")\n for path in [self.wav_path, self.vox2_base_path]:\n # if vox2 directory is not set and vox2 is not a directory \n # we will not process this directory\n if not path or not os.path.exists(path):\n print(f\"{path} is an invalid path, please check again, \"\n \"and we will ignore the vox2 base path\")\n continue\n for file in glob.glob(\n os.path.join(path, \"**\", \"*.wav\"), recursive=True):\n spk = file.split('/wav/')[1].split('/')[0]\n if spk in test_spks:\n continue\n speakers.add(spk)\n audio_files.append(file)\n\n print(\n f\"start to generate the {os.path.join(self.meta_path, 'spk_id2label.txt')}\"\n )\n # encode the train and dev speakers label to spk_id2label.txt\n with open(os.path.join(self.meta_path, 'spk_id2label.txt'), 'w') as f:\n for label, spk_id in enumerate(\n sorted(speakers)): # 1211 vox1, 5994 vox2, 7205 vox1+2\n f.write(f'{spk_id} {label}\\n')\n\n audio_files = sorted(audio_files)\n random.shuffle(audio_files)\n split_idx = int(self.split_ratio * len(audio_files))\n # split_ratio to train\n train_files, dev_files = audio_files[:split_idx], audio_files[\n split_idx:]\n\n self.generate_csv(train_files, os.path.join(self.csv_path, 'train.csv'))\n self.generate_csv(dev_files, os.path.join(self.csv_path, 'dev.csv'))\n\n self.generate_csv(\n enroll_files,\n os.path.join(self.csv_path, 'enroll.csv'),\n split_chunks=False)\n self.generate_csv(\n test_files,\n os.path.join(self.csv_path, 'test.csv'),\n split_chunks=False)\n\n def __getitem__(self, idx):\n return self._convert_to_record(idx)\n\n def __len__(self):\n return len(self._data)\n"
},
{
"path": "audio/paddleaudio/features/__init__.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom .layers import LogMelSpectrogram\nfrom .layers import MelSpectrogram\nfrom .layers import MFCC\nfrom .layers import Spectrogram\n"
},
{
"path": "audio/paddleaudio/features/layers.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom functools import partial\nfrom typing import Optional\nfrom typing import Union\n\nimport paddle\nimport paddle.nn as nn\nfrom paddle import Tensor\n\nfrom ..functional import compute_fbank_matrix\nfrom ..functional import create_dct\nfrom ..functional import power_to_db\nfrom ..functional.window import get_window\n\n__all__ = [\n 'Spectrogram',\n 'MelSpectrogram',\n 'LogMelSpectrogram',\n 'MFCC',\n]\n\n\nclass Spectrogram(nn.Layer):\n \"\"\"Compute spectrogram of given signals, typically audio waveforms.\n The spectrogram is defined as the complex norm of the short-time Fourier transformation.\n\n Args:\n n_fft (int, optional): The number of frequency components of the discrete Fourier transform. Defaults to 512.\n hop_length (Optional[int], optional): The hop length of the short time FFT. If `None`, it is set to `win_length//4`. Defaults to None.\n win_length (Optional[int], optional): The window length of the short time FFT. If `None`, it is set to same as `n_fft`. Defaults to None.\n window (str, optional): The window function applied to the signal before the Fourier transform. Supported window functions: 'hamming', 'hann', 'kaiser', 'gaussian', 'exponential', 'triang', 'bohman', 'blackman', 'cosine', 'tukey', 'taylor'. Defaults to 'hann'.\n power (float, optional): Exponent for the magnitude spectrogram. Defaults to 2.0.\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to 'reflect'.\n dtype (str, optional): Data type of input and window. Defaults to 'float32'.\n \"\"\"\n\n def __init__(self,\n n_fft: int=512,\n hop_length: Optional[int]=None,\n win_length: Optional[int]=None,\n window: str='hann',\n power: float=2.0,\n center: bool=True,\n pad_mode: str='reflect',\n dtype: str='float32') -> None:\n super(Spectrogram, self).__init__()\n\n assert power > 0, 'Power of spectrogram must be > 0.'\n self.power = power\n\n if win_length is None:\n win_length = n_fft\n\n self.fft_window = get_window(\n window, win_length, fftbins=True, dtype=dtype)\n self._stft = partial(\n paddle.signal.stft,\n n_fft=n_fft,\n hop_length=hop_length,\n win_length=win_length,\n window=self.fft_window,\n center=center,\n pad_mode=pad_mode)\n self.register_buffer('fft_window', self.fft_window)\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x (Tensor): Tensor of waveforms with shape `(N, T)`\n\n Returns:\n Tensor: Spectrograms with shape `(N, n_fft//2 + 1, num_frames)`.\n \"\"\"\n stft = self._stft(x)\n spectrogram = paddle.pow(paddle.abs(stft), self.power)\n return spectrogram\n\n\nclass MelSpectrogram(nn.Layer):\n \"\"\"Compute the melspectrogram of given signals, typically audio waveforms. It is computed by multiplying spectrogram with Mel filter bank matrix.\n\n Args:\n sr (int, optional): Sample rate. Defaults to 22050.\n n_fft (int, optional): The number of frequency components of the discrete Fourier transform. Defaults to 512.\n hop_length (Optional[int], optional): The hop length of the short time FFT. If `None`, it is set to `win_length//4`. Defaults to None.\n win_length (Optional[int], optional): The window length of the short time FFT. If `None`, it is set to same as `n_fft`. Defaults to None.\n window (str, optional): The window function applied to the signal before the Fourier transform. Supported window functions: 'hamming', 'hann', 'kaiser', 'gaussian', 'exponential', 'triang', 'bohman', 'blackman', 'cosine', 'tukey', 'taylor'. Defaults to 'hann'.\n power (float, optional): Exponent for the magnitude spectrogram. Defaults to 2.0.\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to 'reflect'.\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n f_min (float, optional): Minimum frequency in Hz. Defaults to 50.0.\n f_max (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n htk (bool, optional): Use HTK formula in computing fbank matrix. Defaults to False.\n norm (Union[str, float], optional): Type of normalization in computing fbank matrix. Slaney-style is used by default. You can specify norm=1.0/2.0 to use customized p-norm normalization. Defaults to 'slaney'.\n dtype (str, optional): Data type of input and window. Defaults to 'float32'.\n \"\"\"\n\n def __init__(self,\n sr: int=22050,\n n_fft: int=512,\n hop_length: Optional[int]=None,\n win_length: Optional[int]=None,\n window: str='hann',\n power: float=2.0,\n center: bool=True,\n pad_mode: str='reflect',\n n_mels: int=64,\n f_min: float=50.0,\n f_max: Optional[float]=None,\n htk: bool=False,\n norm: Union[str, float]='slaney',\n dtype: str='float32') -> None:\n super(MelSpectrogram, self).__init__()\n\n self._spectrogram = Spectrogram(\n n_fft=n_fft,\n hop_length=hop_length,\n win_length=win_length,\n window=window,\n power=power,\n center=center,\n pad_mode=pad_mode,\n dtype=dtype)\n self.n_mels = n_mels\n self.f_min = f_min\n self.f_max = f_max\n self.htk = htk\n self.norm = norm\n if f_max is None:\n f_max = sr // 2\n self.fbank_matrix = compute_fbank_matrix(\n sr=sr,\n n_fft=n_fft,\n n_mels=n_mels,\n f_min=f_min,\n f_max=f_max,\n htk=htk,\n norm=norm,\n dtype=dtype) # float64 for better numerical results\n self.register_buffer('fbank_matrix', self.fbank_matrix)\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x (Tensor): Tensor of waveforms with shape `(N, T)`\n\n Returns:\n Tensor: Mel spectrograms with shape `(N, n_mels, num_frames)`.\n \"\"\"\n spect_feature = self._spectrogram(x)\n mel_feature = paddle.matmul(self.fbank_matrix, spect_feature)\n return mel_feature\n\n\nclass LogMelSpectrogram(nn.Layer):\n \"\"\"Compute log-mel-spectrogram feature of given signals, typically audio waveforms.\n\n Args:\n sr (int, optional): Sample rate. Defaults to 22050.\n n_fft (int, optional): The number of frequency components of the discrete Fourier transform. Defaults to 512.\n hop_length (Optional[int], optional): The hop length of the short time FFT. If `None`, it is set to `win_length//4`. Defaults to None.\n win_length (Optional[int], optional): The window length of the short time FFT. If `None`, it is set to same as `n_fft`. Defaults to None.\n window (str, optional): The window function applied to the signal before the Fourier transform. Supported window functions: 'hamming', 'hann', 'kaiser', 'gaussian', 'exponential', 'triang', 'bohman', 'blackman', 'cosine', 'tukey', 'taylor'. Defaults to 'hann'.\n power (float, optional): Exponent for the magnitude spectrogram. Defaults to 2.0.\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to 'reflect'.\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n f_min (float, optional): Minimum frequency in Hz. Defaults to 50.0.\n f_max (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n htk (bool, optional): Use HTK formula in computing fbank matrix. Defaults to False.\n norm (Union[str, float], optional): Type of normalization in computing fbank matrix. Slaney-style is used by default. You can specify norm=1.0/2.0 to use customized p-norm normalization. Defaults to 'slaney'.\n ref_value (float, optional): The reference value. If smaller than 1.0, the db level of the signal will be pulled up accordingly. Otherwise, the db level is pushed down. Defaults to 1.0.\n amin (float, optional): The minimum value of input magnitude. Defaults to 1e-10.\n top_db (Optional[float], optional): The maximum db value of spectrogram. Defaults to None.\n dtype (str, optional): Data type of input and window. Defaults to 'float32'.\n \"\"\"\n\n def __init__(self,\n sr: int=22050,\n n_fft: int=512,\n hop_length: Optional[int]=None,\n win_length: Optional[int]=None,\n window: str='hann',\n power: float=2.0,\n center: bool=True,\n pad_mode: str='reflect',\n n_mels: int=64,\n f_min: float=50.0,\n f_max: Optional[float]=None,\n htk: bool=False,\n norm: Union[str, float]='slaney',\n ref_value: float=1.0,\n amin: float=1e-10,\n top_db: Optional[float]=None,\n dtype: str='float32') -> None:\n super(LogMelSpectrogram, self).__init__()\n\n self._melspectrogram = MelSpectrogram(\n sr=sr,\n n_fft=n_fft,\n hop_length=hop_length,\n win_length=win_length,\n window=window,\n power=power,\n center=center,\n pad_mode=pad_mode,\n n_mels=n_mels,\n f_min=f_min,\n f_max=f_max,\n htk=htk,\n norm=norm,\n dtype=dtype)\n\n self.ref_value = ref_value\n self.amin = amin\n self.top_db = top_db\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x (Tensor): Tensor of waveforms with shape `(N, T)`\n\n Returns:\n Tensor: Log mel spectrograms with shape `(N, n_mels, num_frames)`.\n \"\"\"\n mel_feature = self._melspectrogram(x)\n log_mel_feature = power_to_db(\n mel_feature,\n ref_value=self.ref_value,\n amin=self.amin,\n top_db=self.top_db)\n return log_mel_feature\n\n\nclass MFCC(nn.Layer):\n \"\"\"Compute mel frequency cepstral coefficients(MFCCs) feature of given waveforms.\n\n Args:\n sr (int, optional): Sample rate. Defaults to 22050.\n n_mfcc (int, optional): [description]. Defaults to 40.\n n_fft (int, optional): The number of frequency components of the discrete Fourier transform. Defaults to 512.\n hop_length (Optional[int], optional): The hop length of the short time FFT. If `None`, it is set to `win_length//4`. Defaults to None.\n win_length (Optional[int], optional): The window length of the short time FFT. If `None`, it is set to same as `n_fft`. Defaults to None.\n window (str, optional): The window function applied to the signal before the Fourier transform. Supported window functions: 'hamming', 'hann', 'kaiser', 'gaussian', 'exponential', 'triang', 'bohman', 'blackman', 'cosine', 'tukey', 'taylor'. Defaults to 'hann'.\n power (float, optional): Exponent for the magnitude spectrogram. Defaults to 2.0.\n center (bool, optional): Whether to pad `x` to make that the :math:`t \\times hop\\\\_length` at the center of `t`-th frame. Defaults to True.\n pad_mode (str, optional): Choose padding pattern when `center` is `True`. Defaults to 'reflect'.\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n f_min (float, optional): Minimum frequency in Hz. Defaults to 50.0.\n f_max (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n htk (bool, optional): Use HTK formula in computing fbank matrix. Defaults to False.\n norm (Union[str, float], optional): Type of normalization in computing fbank matrix. Slaney-style is used by default. You can specify norm=1.0/2.0 to use customized p-norm normalization. Defaults to 'slaney'.\n ref_value (float, optional): The reference value. If smaller than 1.0, the db level of the signal will be pulled up accordingly. Otherwise, the db level is pushed down. Defaults to 1.0.\n amin (float, optional): The minimum value of input magnitude. Defaults to 1e-10.\n top_db (Optional[float], optional): The maximum db value of spectrogram. Defaults to None.\n dtype (str, optional): Data type of input and window. Defaults to 'float32'.\n \"\"\"\n\n def __init__(self,\n sr: int=22050,\n n_mfcc: int=40,\n n_fft: int=512,\n hop_length: Optional[int]=None,\n win_length: Optional[int]=None,\n window: str='hann',\n power: float=2.0,\n center: bool=True,\n pad_mode: str='reflect',\n n_mels: int=64,\n f_min: float=50.0,\n f_max: Optional[float]=None,\n htk: bool=False,\n norm: Union[str, float]='slaney',\n ref_value: float=1.0,\n amin: float=1e-10,\n top_db: Optional[float]=None,\n dtype: str=paddle.float32) -> None:\n super(MFCC, self).__init__()\n assert n_mfcc <= n_mels, 'n_mfcc cannot be larger than n_mels: %d vs %d' % (\n n_mfcc, n_mels)\n self._log_melspectrogram = LogMelSpectrogram(\n sr=sr,\n n_fft=n_fft,\n hop_length=hop_length,\n win_length=win_length,\n window=window,\n power=power,\n center=center,\n pad_mode=pad_mode,\n n_mels=n_mels,\n f_min=f_min,\n f_max=f_max,\n htk=htk,\n norm=norm,\n ref_value=ref_value,\n amin=amin,\n top_db=top_db,\n dtype=dtype)\n self.dct_matrix = create_dct(n_mfcc=n_mfcc, n_mels=n_mels, dtype=dtype)\n self.register_buffer('dct_matrix', self.dct_matrix)\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x (Tensor): Tensor of waveforms with shape `(N, T)`\n\n Returns:\n Tensor: Mel frequency cepstral coefficients with shape `(N, n_mfcc, num_frames)`.\n \"\"\"\n log_mel_feature = self._log_melspectrogram(x)\n mfcc = paddle.matmul(\n log_mel_feature.transpose((0, 2, 1)), self.dct_matrix).transpose(\n (0, 2, 1)) # (B, n_mels, L)\n return mfcc\n"
},
{
"path": "audio/paddleaudio/functional/__init__.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom .functional import compute_fbank_matrix\nfrom .functional import create_dct\nfrom .functional import fft_frequencies\nfrom .functional import hz_to_mel\nfrom .functional import mel_frequencies\nfrom .functional import mel_to_hz\nfrom .functional import power_to_db\n"
},
{
"path": "audio/paddleaudio/functional/functional.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n# Modified from librosa(https://github.com/librosa/librosa)\nimport math\nfrom typing import Optional\nfrom typing import Union\n\nimport paddle\nfrom paddle import Tensor\n\n__all__ = [\n 'hz_to_mel',\n 'mel_to_hz',\n 'mel_frequencies',\n 'fft_frequencies',\n 'compute_fbank_matrix',\n 'power_to_db',\n 'create_dct',\n]\n\n\ndef hz_to_mel(freq: Union[Tensor, float],\n htk: bool=False) -> Union[Tensor, float]:\n \"\"\"Convert Hz to Mels.\n\n Args:\n freq (Union[Tensor, float]): The input tensor with arbitrary shape.\n htk (bool, optional): Use htk scaling. Defaults to False.\n\n Returns:\n Union[Tensor, float]: Frequency in mels.\n \"\"\"\n\n if htk:\n if isinstance(freq, Tensor):\n return 2595.0 * paddle.log10(1.0 + freq / 700.0)\n else:\n return 2595.0 * math.log10(1.0 + freq / 700.0)\n\n # Fill in the linear part\n f_min = 0.0\n f_sp = 200.0 / 3\n\n mels = (freq - f_min) / f_sp\n\n # Fill in the log-scale part\n\n min_log_hz = 1000.0 # beginning of log region (Hz)\n min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)\n logstep = math.log(6.4) / 27.0 # step size for log region\n\n if isinstance(freq, Tensor):\n target = min_log_mel + paddle.log(\n freq / min_log_hz + 1e-10) / logstep # prevent nan with 1e-10\n mask = (freq > min_log_hz).astype(freq.dtype)\n mels = target * mask + mels * (\n 1 - mask) # will replace by masked_fill OP in future\n else:\n if freq >= min_log_hz:\n mels = min_log_mel + math.log(freq / min_log_hz + 1e-10) / logstep\n\n return mels\n\n\ndef mel_to_hz(mel: Union[float, Tensor],\n htk: bool=False) -> Union[float, Tensor]:\n \"\"\"Convert mel bin numbers to frequencies.\n\n Args:\n mel (Union[float, Tensor]): The mel frequency represented as a tensor with arbitrary shape.\n htk (bool, optional): Use htk scaling. Defaults to False.\n\n Returns:\n Union[float, Tensor]: Frequencies in Hz.\n \"\"\"\n if htk:\n return 700.0 * (10.0**(mel / 2595.0) - 1.0)\n\n f_min = 0.0\n f_sp = 200.0 / 3\n freqs = f_min + f_sp * mel\n # And now the nonlinear scale\n min_log_hz = 1000.0 # beginning of log region (Hz)\n min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)\n logstep = math.log(6.4) / 27.0 # step size for log region\n if isinstance(mel, Tensor):\n target = min_log_hz * paddle.exp(logstep * (mel - min_log_mel))\n mask = (mel > min_log_mel).astype(mel.dtype)\n freqs = target * mask + freqs * (\n 1 - mask) # will replace by masked_fill OP in future\n else:\n if mel >= min_log_mel:\n freqs = min_log_hz * math.exp(logstep * (mel - min_log_mel))\n\n return freqs\n\n\ndef mel_frequencies(n_mels: int=64,\n f_min: float=0.0,\n f_max: float=11025.0,\n htk: bool=False,\n dtype: str='float32') -> Tensor:\n \"\"\"Compute mel frequencies.\n\n Args:\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n f_min (float, optional): Minimum frequency in Hz. Defaults to 0.0.\n fmax (float, optional): Maximum frequency in Hz. Defaults to 11025.0.\n htk (bool, optional): Use htk scaling. Defaults to False.\n dtype (str, optional): The data type of the return frequencies. Defaults to 'float32'.\n\n Returns:\n Tensor: Tensor of n_mels frequencies in Hz with shape `(n_mels,)`.\n \"\"\"\n # 'Center freqs' of mel bands - uniformly spaced between limits\n min_mel = hz_to_mel(f_min, htk=htk)\n max_mel = hz_to_mel(f_max, htk=htk)\n mels = paddle.linspace(min_mel, max_mel, n_mels, dtype=dtype)\n freqs = mel_to_hz(mels, htk=htk)\n return freqs\n\n\ndef fft_frequencies(sr: int, n_fft: int, dtype: str='float32') -> Tensor:\n \"\"\"Compute fourier frequencies.\n\n Args:\n sr (int): Sample rate.\n n_fft (int): Number of fft bins.\n dtype (str, optional): The data type of the return frequencies. Defaults to 'float32'.\n\n Returns:\n Tensor: FFT frequencies in Hz with shape `(n_fft//2 + 1,)`.\n \"\"\"\n return paddle.linspace(0, float(sr) / 2, int(1 + n_fft // 2), dtype=dtype)\n\n\ndef compute_fbank_matrix(sr: int,\n n_fft: int,\n n_mels: int=64,\n f_min: float=0.0,\n f_max: Optional[float]=None,\n htk: bool=False,\n norm: Union[str, float]='slaney',\n dtype: str='float32') -> Tensor:\n \"\"\"Compute fbank matrix.\n\n Args:\n sr (int): Sample rate.\n n_fft (int): Number of fft bins.\n n_mels (int, optional): Number of mel bins. Defaults to 64.\n f_min (float, optional): Minimum frequency in Hz. Defaults to 0.0.\n f_max (Optional[float], optional): Maximum frequency in Hz. Defaults to None.\n htk (bool, optional): Use htk scaling. Defaults to False.\n norm (Union[str, float], optional): Type of normalization. Defaults to 'slaney'.\n dtype (str, optional): The data type of the return matrix. Defaults to 'float32'.\n\n Returns:\n Tensor: Mel transform matrix with shape `(n_mels, n_fft//2 + 1)`.\n \"\"\"\n\n if f_max is None:\n f_max = float(sr) / 2\n\n # Initialize the weights\n weights = paddle.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)\n\n # Center freqs of each FFT bin\n fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft, dtype=dtype)\n\n # 'Center freqs' of mel bands - uniformly spaced between limits\n mel_f = mel_frequencies(\n n_mels + 2, f_min=f_min, f_max=f_max, htk=htk, dtype=dtype)\n\n fdiff = mel_f[1:] - mel_f[:-1] #np.diff(mel_f)\n ramps = mel_f.unsqueeze(1) - fftfreqs.unsqueeze(0)\n #ramps = np.subtract.outer(mel_f, fftfreqs)\n\n for i in range(n_mels):\n # lower and upper slopes for all bins\n lower = -ramps[i] / fdiff[i]\n upper = ramps[i + 2] / fdiff[i + 1]\n\n # .. then intersect them with each other and zero\n weights[i] = paddle.maximum(\n paddle.zeros_like(lower), paddle.minimum(lower, upper))\n\n # Slaney-style mel is scaled to be approx constant energy per channel\n if norm == 'slaney':\n enorm = 2.0 / (mel_f[2:n_mels + 2] - mel_f[:n_mels])\n weights *= enorm.unsqueeze(1)\n elif isinstance(norm, int) or isinstance(norm, float):\n weights = paddle.nn.functional.normalize(weights, p=norm, axis=-1)\n\n return weights\n\n\ndef power_to_db(spect: Tensor,\n ref_value: float=1.0,\n amin: float=1e-10,\n top_db: Optional[float]=None) -> Tensor:\n \"\"\"Convert a power spectrogram (amplitude squared) to decibel (dB) units. The function computes the scaling `10 * log10(x / ref)` in a numerically stable way.\n\n Args:\n spect (Tensor): STFT power spectrogram.\n ref_value (float, optional): The reference value. If smaller than 1.0, the db level of the signal will be pulled up accordingly. Otherwise, the db level is pushed down. Defaults to 1.0.\n amin (float, optional): Minimum threshold. Defaults to 1e-10.\n top_db (Optional[float], optional): Threshold the output at `top_db` below the peak. Defaults to None.\n\n Returns:\n Tensor: Power spectrogram in db scale.\n \"\"\"\n if amin <= 0:\n raise Exception(\"amin must be strictly positive\")\n\n if ref_value <= 0:\n raise Exception(\"ref_value must be strictly positive\")\n\n ones = paddle.ones_like(spect)\n log_spec = 10.0 * paddle.log10(paddle.maximum(ones * amin, spect))\n log_spec -= 10.0 * math.log10(max(ref_value, amin))\n\n if top_db is not None:\n if top_db < 0:\n raise Exception(\"top_db must be non-negative\")\n log_spec = paddle.maximum(log_spec, ones * (log_spec.max() - top_db))\n\n return log_spec\n\n\ndef create_dct(n_mfcc: int,\n n_mels: int,\n norm: Optional[str]='ortho',\n dtype: str='float32') -> Tensor:\n \"\"\"Create a discrete cosine transform(DCT) matrix.\n\n Args:\n n_mfcc (int): Number of mel frequency cepstral coefficients. \n n_mels (int): Number of mel filterbanks.\n norm (Optional[str], optional): Normalization type. Defaults to 'ortho'.\n dtype (str, optional): The data type of the return matrix. Defaults to 'float32'.\n\n Returns:\n Tensor: The DCT matrix with shape `(n_mels, n_mfcc)`.\n \"\"\"\n n = paddle.arange(n_mels, dtype=dtype)\n k = paddle.arange(n_mfcc, dtype=dtype).unsqueeze(1)\n dct = paddle.cos(math.pi / float(n_mels) * (n + 0.5) *\n k) # size (n_mfcc, n_mels)\n if norm is None:\n dct *= 2.0\n else:\n assert norm == \"ortho\"\n dct[0] *= 1.0 / math.sqrt(2.0)\n dct *= math.sqrt(2.0 / float(n_mels))\n return dct.T\n"
},
{
"path": "audio/paddleaudio/functional/window.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\nimport math\nfrom typing import List\nfrom typing import Tuple\nfrom typing import Union\n\nimport paddle\nfrom paddle import Tensor\n\n\nclass WindowFunctionRegister(object):\n def __init__(self):\n self._functions_dict = dict()\n\n def register(self):\n def add_subfunction(func):\n name = func.__name__\n self._functions_dict[name] = func\n return func\n\n return add_subfunction\n\n def get(self, name):\n return self._functions_dict[name]\n\n\nwindow_function_register = WindowFunctionRegister()\n\n\n@window_function_register.register()\ndef _cat(x: List[Tensor], data_type: str) -> Tensor:\n l = [paddle.to_tensor(_, data_type) for _ in x]\n return paddle.concat(l)\n\n\n@window_function_register.register()\ndef _acosh(x: Union[Tensor, float]) -> Tensor:\n if isinstance(x, float):\n return math.log(x + math.sqrt(x**2 - 1))\n return paddle.log(x + paddle.sqrt(paddle.square(x) - 1))\n\n\n@window_function_register.register()\ndef _extend(M: int, sym: bool) -> bool:\n \"\"\"Extend window by 1 sample if needed for DFT-even symmetry.\"\"\"\n if not sym:\n return M + 1, True\n else:\n return M, False\n\n\n@window_function_register.register()\ndef _len_guards(M: int) -> bool:\n \"\"\"Handle small or incorrect window lengths.\"\"\"\n if int(M) != M or M < 0:\n raise ValueError('Window length M must be a non-negative integer')\n\n return M <= 1\n\n\n@window_function_register.register()\ndef _truncate(w: Tensor, needed: bool) -> Tensor:\n \"\"\"Truncate window by 1 sample if needed for DFT-even symmetry.\"\"\"\n if needed:\n return w[:-1]\n else:\n return w\n\n\n@window_function_register.register()\ndef _general_gaussian(\n M: int, p, sig, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a window with a generalized Gaussian shape.\n This function is consistent with scipy.signal.windows.general_gaussian().\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n\n n = paddle.arange(0, M, dtype=dtype) - (M - 1.0) / 2.0\n w = paddle.exp(-0.5 * paddle.abs(n / sig) ** (2 * p))\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _general_cosine(\n M: int, a: float, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a generic weighted sum of cosine terms window.\n This function is consistent with scipy.signal.windows.general_cosine().\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n fac = paddle.linspace(-math.pi, math.pi, M, dtype=dtype)\n w = paddle.zeros((M,), dtype=dtype)\n for k in range(len(a)):\n w += a[k] * paddle.cos(k * fac)\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _general_hamming(\n M: int, alpha: float, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a generalized Hamming window.\n This function is consistent with scipy.signal.windows.general_hamming()\n \"\"\"\n return _general_cosine(M, [alpha, 1.0 - alpha], sym, dtype=dtype)\n\n\n@window_function_register.register()\ndef _taylor(\n M: int, nbar=4, sll=30, norm=True, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a Taylor window.\n The Taylor window taper function approximates the Dolph-Chebyshev window's\n constant sidelobe level for a parameterized number of near-in sidelobes.\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n # Original text uses a negative sidelobe level parameter and then negates\n # it in the calculation of B. To keep consistent with other methods we\n # assume the sidelobe level parameter to be positive.\n B = 10 ** (sll / 20)\n A = _acosh(B) / math.pi\n s2 = nbar**2 / (A**2 + (nbar - 0.5) ** 2)\n ma = paddle.arange(1, nbar, dtype=dtype)\n\n Fm = paddle.empty((nbar - 1,), dtype=dtype)\n signs = paddle.empty_like(ma)\n signs[::2] = 1\n signs[1::2] = -1\n m2 = ma * ma\n for mi in range(len(ma)):\n numer = signs[mi] * paddle.prod(\n 1 - m2[mi] / s2 / (A**2 + (ma - 0.5) ** 2)\n )\n if mi == 0:\n denom = 2 * paddle.prod(1 - m2[mi] / m2[mi + 1 :])\n elif mi == len(ma) - 1:\n denom = 2 * paddle.prod(1 - m2[mi] / m2[:mi])\n else:\n denom = (\n 2\n * paddle.prod(1 - m2[mi] / m2[:mi])\n * paddle.prod(1 - m2[mi] / m2[mi + 1 :])\n )\n\n Fm[mi] = numer / denom\n\n def W(n):\n return 1 + 2 * paddle.matmul(\n Fm.unsqueeze(0),\n paddle.cos(2 * math.pi * ma.unsqueeze(1) * (n - M / 2.0 + 0.5) / M),\n )\n\n w = W(paddle.arange(0, M, dtype=dtype))\n\n # normalize (Note that this is not described in the original text [1])\n if norm:\n scale = 1.0 / W((M - 1) / 2)\n w *= scale\n w = w.squeeze()\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _hamming(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a Hamming window.\n The Hamming window is a taper formed by using a raised cosine with\n non-zero endpoints, optimized to minimize the nearest side lobe.\n \"\"\"\n return _general_hamming(M, 0.54, sym, dtype=dtype)\n\n\n@window_function_register.register()\ndef _hann(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a Hann window.\n The Hann window is a taper formed by using a raised cosine or sine-squared\n with ends that touch zero.\n \"\"\"\n return _general_hamming(M, 0.5, sym, dtype=dtype)\n\n\n@window_function_register.register()\ndef _tukey(\n M: int, alpha=0.5, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a Tukey window.\n The Tukey window is also known as a tapered cosine window.\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n\n if alpha <= 0:\n return paddle.ones((M,), dtype=dtype)\n elif alpha >= 1.0:\n return hann(M, sym=sym)\n\n M, needs_trunc = _extend(M, sym)\n\n n = paddle.arange(0, M, dtype=dtype)\n width = int(alpha * (M - 1) / 2.0)\n n1 = n[0 : width + 1]\n n2 = n[width + 1 : M - width - 1]\n n3 = n[M - width - 1 :]\n\n w1 = 0.5 * (1 + paddle.cos(math.pi * (-1 + 2.0 * n1 / alpha / (M - 1))))\n w2 = paddle.ones(n2.shape, dtype=dtype)\n w3 = 0.5 * (\n 1\n + paddle.cos(math.pi * (-2.0 / alpha + 1 + 2.0 * n3 / alpha / (M - 1)))\n )\n w = paddle.concat([w1, w2, w3])\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _gaussian(\n M: int, std: float, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute a Gaussian window.\n The Gaussian widows has a Gaussian shape defined by the standard deviation(std).\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n\n n = paddle.arange(0, M, dtype=dtype) - (M - 1.0) / 2.0\n sig2 = 2 * std * std\n w = paddle.exp(-(n**2) / sig2)\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _exponential(\n M: int, center=None, tau=1.0, sym: bool = True, dtype: str = 'float64'\n) -> Tensor:\n \"\"\"Compute an exponential (or Poisson) window.\"\"\"\n if sym and center is not None:\n raise ValueError(\"If sym==True, center must be None.\")\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n\n if center is None:\n center = (M - 1) / 2\n\n n = paddle.arange(0, M, dtype=dtype)\n w = paddle.exp(-paddle.abs(n - center) / tau)\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _triang(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a triangular window.\"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n\n n = paddle.arange(1, (M + 1) // 2 + 1, dtype=dtype)\n if M % 2 == 0:\n w = (2 * n - 1.0) / M\n w = paddle.concat([w, w[::-1]])\n else:\n w = 2 * n / (M + 1.0)\n w = paddle.concat([w, w[-2::-1]])\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _bohman(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a Bohman window.\n The Bohman window is the autocorrelation of a cosine window.\n \"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n\n fac = paddle.abs(paddle.linspace(-1, 1, M, dtype=dtype)[1:-1])\n w = (1 - fac) * paddle.cos(math.pi * fac) + 1.0 / math.pi * paddle.sin(\n math.pi * fac\n )\n w = _cat([0, w, 0], dtype)\n\n return _truncate(w, needs_trunc)\n\n\n@window_function_register.register()\ndef _blackman(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a Blackman window.\n The Blackman window is a taper formed by using the first three terms of\n a summation of cosines. It was designed to have close to the minimal\n leakage possible. It is close to optimal, only slightly worse than a\n Kaiser window.\n \"\"\"\n return _general_cosine(M, [0.42, 0.50, 0.08], sym, dtype=dtype)\n\n\n@window_function_register.register()\ndef _cosine(M: int, sym: bool = True, dtype: str = 'float64') -> Tensor:\n \"\"\"Compute a window with a simple cosine shape.\"\"\"\n if _len_guards(M):\n return paddle.ones((M,), dtype=dtype)\n M, needs_trunc = _extend(M, sym)\n w = paddle.sin(math.pi / M * (paddle.arange(0, M, dtype=dtype) + 0.5))\n\n return _truncate(w, needs_trunc)\n\n\ndef get_window(\n window: Union[str, Tuple[str, float]],\n win_length: int,\n fftbins: bool = True,\n dtype: str = 'float64',\n) -> Tensor:\n \"\"\"Return a window of a given length and type.\n\n Args:\n window (Union[str, Tuple[str, float]]): The window function applied to the signal before the Fourier transform. Supported window functions: 'hamming', 'hann', 'gaussian', 'general_gaussian', 'exponential', 'triang', 'bohman', 'blackman', 'cosine', 'tukey', 'taylor'.\n win_length (int): Number of samples.\n fftbins (bool, optional): If True, create a \"periodic\" window. Otherwise, create a \"symmetric\" window, for use in filter design. Defaults to True.\n dtype (str, optional): The data type of the return window. Defaults to 'float64'.\n\n Returns:\n Tensor: The window represented as a tensor.\n\n Examples:\n .. code-block:: python\n\n import paddle\n\n n_fft = 512\n cosine_window = paddle.audio.functional.get_window('cosine', n_fft)\n\n std = 7\n gaussian_window = paddle.audio.functional.get_window(('gaussian',std), n_fft)\n \"\"\"\n sym = not fftbins\n\n args = ()\n if isinstance(window, tuple):\n winstr = window[0]\n if len(window) > 1:\n args = window[1:]\n elif isinstance(window, str):\n if window in ['gaussian', 'exponential']:\n raise ValueError(\n \"The '\" + window + \"' window needs one or \"\n \"more parameters -- pass a tuple.\"\n )\n else:\n winstr = window\n else:\n raise ValueError(\n \"%s as window type is not supported.\" % str(type(window))\n )\n\n try:\n winfunc = window_function_register.get('_' + winstr)\n except KeyError as e:\n raise ValueError(\"Unknown window type.\") from e\n\n params = (win_length,) + args\n kwargs = {'sym': sym}\n return winfunc(*params, dtype=dtype, **kwargs)\n"
},
{
"path": "audio/paddleaudio/kaldi/__init__.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom .kaldi import fbank\n#from .kaldi import pitch\n"
},
{
"path": "audio/paddleaudio/kaldi/kaldi.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nimport paddleaudio\nfrom paddleaudio._internal import module_utils\n\n__all__ = [\n 'fbank',\n]\n\n\n@module_utils.requires_kaldi()\ndef fbank(\n wav,\n samp_freq: int=16000,\n frame_shift_ms: float=10.0,\n frame_length_ms: float=25.0,\n dither: float=0.0,\n preemph_coeff: float=0.97,\n remove_dc_offset: bool=True,\n window_type: str='povey',\n round_to_power_of_two: bool=True,\n blackman_coeff: float=0.42,\n snip_edges: bool=True,\n max_feature_vectors: int=-1,\n num_bins: int=23,\n low_freq: float=20,\n high_freq: float=0,\n vtln_low: float=100,\n vtln_high: float=-500,\n debug_mel: bool=False,\n htk_mode: bool=False,\n use_energy: bool=False, # fbank opts\n energy_floor: float=0.0,\n raw_energy: bool=True,\n htk_compat: bool=False,\n use_log_fbank: bool=True,\n use_power: bool=True):\n frame_opts = paddleaudio._paddleaudio.FrameExtractionOptions()\n mel_opts = paddleaudio._paddleaudio.MelBanksOptions()\n fbank_opts = paddleaudio._paddleaudio.FbankOptions()\n frame_opts.samp_freq = samp_freq\n frame_opts.frame_shift_ms = frame_shift_ms\n frame_opts.frame_length_ms = frame_length_ms\n frame_opts.dither = dither\n frame_opts.preemph_coeff = preemph_coeff\n frame_opts.remove_dc_offset = remove_dc_offset\n frame_opts.window_type = window_type\n frame_opts.round_to_power_of_two = round_to_power_of_two\n frame_opts.blackman_coeff = blackman_coeff\n frame_opts.snip_edges = snip_edges\n frame_opts.max_feature_vectors = max_feature_vectors\n\n mel_opts.num_bins = num_bins\n mel_opts.low_freq = low_freq\n mel_opts.high_freq = high_freq\n mel_opts.vtln_low = vtln_low\n mel_opts.vtln_high = vtln_high\n mel_opts.debug_mel = debug_mel\n mel_opts.htk_mode = htk_mode\n\n fbank_opts.use_energy = use_energy\n fbank_opts.energy_floor = energy_floor\n fbank_opts.raw_energy = raw_energy\n fbank_opts.htk_compat = htk_compat\n fbank_opts.use_log_fbank = use_log_fbank\n fbank_opts.use_power = use_power\n feat = paddleaudio._paddleaudio.ComputeFbank(frame_opts, mel_opts,\n fbank_opts, wav)\n return feat\n\n\n#@module_utils.requires_kaldi()\n#def pitch(wav,\n#samp_freq: int=16000,\n#frame_shift_ms: float=10.0,\n#frame_length_ms: float=25.0,\n#preemph_coeff: float=0.0,\n#min_f0: int=50,\n#max_f0: int=400,\n#soft_min_f0: float=10.0,\n#penalty_factor: float=0.1,\n#lowpass_cutoff: int=1000,\n#resample_freq: int=4000,\n#delta_pitch: float=0.005,\n#nccf_ballast: int=7000,\n#lowpass_filter_width: int=1,\n#upsample_filter_width: int=5,\n#max_frames_latency: int=0,\n#frames_per_chunk: int=0,\n#simulate_first_pass_online: bool=False,\n#recompute_frame: int=500,\n#nccf_ballast_online: bool=False,\n#snip_edges: bool=True):\n#pitch_opts = paddleaudio._paddleaudio.PitchExtractionOptions()\n#pitch_opts.samp_freq = samp_freq\n#pitch_opts.frame_shift_ms = frame_shift_ms\n#pitch_opts.frame_length_ms = frame_length_ms\n#pitch_opts.preemph_coeff = preemph_coeff\n#pitch_opts.min_f0 = min_f0\n#pitch_opts.max_f0 = max_f0\n#pitch_opts.soft_min_f0 = soft_min_f0\n#pitch_opts.penalty_factor = penalty_factor\n#pitch_opts.lowpass_cutoff = lowpass_cutoff\n#pitch_opts.resample_freq = resample_freq\n#pitch_opts.delta_pitch = delta_pitch\n#pitch_opts.nccf_ballast = nccf_ballast\n#pitch_opts.lowpass_filter_width = lowpass_filter_width\n#pitch_opts.upsample_filter_width = upsample_filter_width\n#pitch_opts.max_frames_latency = max_frames_latency\n#pitch_opts.frames_per_chunk = frames_per_chunk\n#pitch_opts.simulate_first_pass_online = simulate_first_pass_online\n#pitch_opts.recompute_frame = recompute_frame\n#pitch_opts.nccf_ballast_online = nccf_ballast_online\n#pitch_opts.snip_edges = snip_edges\n#pitch = paddleaudio._paddleaudio.ComputeKaldiPitch(pitch_opts, wav)\n#return pitch\n"
},
{
"path": "audio/paddleaudio/metric/__init__.py",
"content": "# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom .eer import compute_eer\nfrom .eer import compute_minDCF\n"
},
{
"path": "audio/paddleaudio/metric/eer.py",
"content": "# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\nfrom typing import List\n\nimport numpy as np\nimport paddle\nfrom sklearn.metrics import roc_curve\n\n\ndef compute_eer(labels: np.ndarray, scores: np.ndarray) -> List[float]:\n \"\"\"Compute EER and return score threshold.\n\n Args:\n labels (np.ndarray): the trial label, shape: [N], one-dimension, N refer to the samples num\n scores (np.ndarray): the trial scores, shape: [N], one-dimension, N refer to the samples num\n\n Returns:\n List[float]: eer and the specific threshold\n \"\"\"\n fpr, tpr, threshold = roc_curve(y_true=labels, y_score=scores)\n fnr = 1 - tpr\n eer_threshold = threshold[np.nanargmin(np.absolute((fnr - fpr)))]\n eer = fpr[np.nanargmin(np.absolute((fnr - fpr)))]\n return eer, eer_threshold\n\n\ndef compute_minDCF(positive_scores,\n negative_scores,\n c_miss=1.0,\n c_fa=1.0,\n p_target=0.01):\n \"\"\"\n This is modified from SpeechBrain\n https://github.com/speechbrain/speechbrain/blob/085be635c07f16d42cd1295045bc46c407f1e15b/speechbrain/utils/metric_stats.py#L509\n Computes the minDCF metric normally used to evaluate speaker verification\n systems. The min_DCF is the minimum of the following C_det function computed\n within the defined threshold range:\n\n C_det = c_miss * p_miss * p_target + c_fa * p_fa * (1 -p_target)\n\n where p_miss is the missing probability and p_fa is the probability of having\n a false alarm.\n\n Args:\n positive_scores (Paddle.Tensor): The scores from entries of the same class.\n negative_scores (Paddle.Tensor): The scores from entries of different classes.\n c_miss (float, optional): Cost assigned to a missing error (default 1.0).\n c_fa (float, optional): Cost assigned to a false alarm (default 1.0).\n p_target (float, optional): Prior probability of having a target (default 0.01).\n\n Returns:\n List[float]: min dcf and the specific threshold\n \"\"\"\n # Computing candidate thresholds\n if len(positive_scores.shape) > 1:\n positive_scores = positive_scores.squeeze()\n\n if len(negative_scores.shape) > 1:\n negative_scores = negative_scores.squeeze()\n\n thresholds = paddle.sort(paddle.concat([positive_scores, negative_scores]))\n thresholds = paddle.unique(thresholds)\n\n # Adding intermediate thresholds\n interm_thresholds = (thresholds[0:-1] + thresholds[1:]) / 2\n thresholds = paddle.sort(paddle.concat([thresholds, interm_thresholds]))\n\n # Computing False Rejection Rate (miss detection)\n positive_scores = paddle.concat(\n len(thresholds) * [positive_scores.unsqueeze(0)])\n pos_scores_threshold = positive_scores.transpose(perm=[1, 0]) <= thresholds\n p_miss = (pos_scores_threshold.sum(0)\n ).astype(\"float32\") / positive_scores.shape[1]\n del positive_scores\n del pos_scores_threshold\n\n # Computing False Acceptance Rate (false alarm)\n negative_scores = paddle.concat(\n len(thresholds) * [negative_scores.unsqueeze(0)])\n neg_scores_threshold = negative_scores.transpose(perm=[1, 0]) > thresholds\n p_fa = (neg_scores_threshold.sum(0)\n ).astype(\"float32\") / negative_scores.shape[1]\n del negative_scores\n del neg_scores_threshold\n\n c_det = c_miss * p_miss * p_target + c_fa * p_fa * (1 - p_target)\n c_min = paddle.min(c_det, axis=0)\n min_index = paddle.argmin(c_det, axis=0)\n return float(c_min), float(thresholds[min_index])\n"
},
{
"path": "audio/paddleaudio/sox_effects/__init__.py",
"content": "from paddleaudio._internal import module_utils as _mod_utils\n\nfrom .sox_effects import apply_effects_file\nfrom .sox_effects import apply_effects_tensor\nfrom .sox_effects import effect_names\nfrom .sox_effects import init_sox_effects\nfrom .sox_effects import shutdown_sox_effects\n\nif _mod_utils.is_sox_available():\n import atexit\n\n init_sox_effects()\n atexit.register(shutdown_sox_effects)\n\n__all__ = [\n \"init_sox_effects\",\n \"shutdown_sox_effects\",\n \"effect_names\",\n \"apply_effects_tensor\",\n \"apply_effects_file\",\n]\n"
},
{
"path": "audio/paddleaudio/sox_effects/sox_effects.py",
"content": "import os\nfrom typing import List\nfrom typing import Optional\nfrom typing import Tuple\n\nimport paddle\nimport paddleaudio\nfrom paddleaudio._internal import module_utils as _mod_utils\nfrom paddleaudio.utils.sox_utils import list_effects\n\n#code is from: https://github.com/pytorch/audio/blob/main/torchaudio/sox_effects/sox_effects.py\n\n\n@_mod_utils.requires_sox()\ndef init_sox_effects():\n \"\"\"Initialize resources required to use sox effects.\n\n Note:\n You do not need to call this function manually. It is called automatically.\n\n Once initialized, you do not need to call this function again across the multiple uses of\n sox effects though it is safe to do so as long as :func:`shutdown_sox_effects` is not called yet.\n Once :func:`shutdown_sox_effects` is called, you can no longer use SoX effects and initializing\n again will result in error.\n \"\"\"\n paddleaudio._paddleaudio.sox_effects_initialize_sox_effects()\n\n\n@_mod_utils.requires_sox()\ndef shutdown_sox_effects():\n \"\"\"Clean up resources required to use sox effects.\n\n Note:\n You do not need to call this function manually. It is called automatically.\n\n It is safe to call this function multiple times.\n Once :py:func:`shutdown_sox_effects` is called, you can no longer use SoX effects and\n initializing again will result in error.\n \"\"\"\n paddleaudio._paddleaudio.sox_effects_shutdown_sox_effects()\n\n\n@_mod_utils.requires_sox()\ndef effect_names() -> List[str]:\n \"\"\"Gets list of valid sox effect names\n\n Returns:\n List[str]: list of available effect names.\n\n Example\n >>> paddleaudio.sox_effects.effect_names()\n ['allpass', 'band', 'bandpass', ... ]\n \"\"\"\n return list(list_effects().keys())\n\n\n@_mod_utils.requires_sox()\ndef apply_effects_tensor(\n tensor: paddle.Tensor,\n sample_rate: int,\n effects: List[List[str]],\n channels_first: bool=True, ) -> Tuple[paddle.Tensor, int]:\n \"\"\"Apply sox effects to given Tensor\n\n .. devices:: CPU\n\n Note:\n This function only works on CPU Tensors.\n This function works in the way very similar to ``sox`` command, however there are slight\n differences. For example, ``sox`` command adds certain effects automatically (such as\n ``rate`` effect after ``speed`` and ``pitch`` and other effects), but this function does\n only applies the given effects. (Therefore, to actually apply ``speed`` effect, you also\n need to give ``rate`` effect with desired sampling rate.).\n\n Args:\n tensor (paddle.Tensor): Input 2D CPU Tensor.\n sample_rate (int): Sample rate\n effects (List[List[str]]): List of effects.\n channels_first (bool, optional): Indicates if the input Tensor's dimension is\n `[channels, time]` or `[time, channels]`\n\n Returns:\n (Tensor, int): Resulting Tensor and sample rate.\n The resulting Tensor has the same ``dtype`` as the input Tensor, and\n the same channels order. The shape of the Tensor can be different based on the\n effects applied. Sample rate can also be different based on the effects applied.\n\n Example - Basic usage\n >>>\n >>> # Defines the effects to apply\n >>> effects = [\n ... ['gain', '-n'], # normalises to 0dB\n ... ['pitch', '5'], # 5 cent pitch shift\n ... ['rate', '8000'], # resample to 8000 Hz\n ... ]\n >>>\n >>> # Generate pseudo wave:\n >>> # normalized, channels first, 2ch, sampling rate 16000, 1 second\n >>> sample_rate = 16000\n >>> waveform = 2 * paddle.rand([2, sample_rate * 1]) - 1\n >>> waveform.shape\n paddle.Size([2, 16000])\n >>> waveform\n tensor([[ 0.3138, 0.7620, -0.9019, ..., -0.7495, -0.4935, 0.5442],\n [-0.0832, 0.0061, 0.8233, ..., -0.5176, -0.9140, -0.2434]])\n >>>\n >>> # Apply effects\n >>> waveform, sample_rate = apply_effects_tensor(\n ... wave_form, sample_rate, effects, channels_first=True)\n >>>\n >>> # Check the result\n >>> # The new waveform is sampling rate 8000, 1 second.\n >>> # normalization and channel order are preserved\n >>> waveform.shape\n paddle.Size([2, 8000])\n >>> waveform\n tensor([[ 0.5054, -0.5518, -0.4800, ..., -0.0076, 0.0096, -0.0110],\n [ 0.1331, 0.0436, -0.3783, ..., -0.0035, 0.0012, 0.0008]])\n >>> sample_rate\n 8000\n\n \"\"\"\n tensor_np = tensor.numpy()\n ret = paddleaudio._paddleaudio.sox_effects_apply_effects_tensor(\n tensor_np, sample_rate, effects, channels_first)\n if ret is not None:\n return (paddle.to_tensor(ret[0]), ret[1])\n raise RuntimeError(\"Failed to apply sox effect\")\n\n\n@_mod_utils.requires_sox()\ndef apply_effects_file(\n path: str,\n effects: List[List[str]],\n normalize: bool=True,\n channels_first: bool=True,\n format: Optional[str]=None, ) -> Tuple[paddle.Tensor, int]:\n \"\"\"Apply sox effects to the audio file and load the resulting data as Tensor\n\n Note:\n This function works in the way very similar to ``sox`` command, however there are slight\n differences. For example, ``sox`` command adds certain effects automatically (such as\n ``rate`` effect after ``speed``, ``pitch`` etc), but this function only applies the given\n effects. Therefore, to actually apply ``speed`` effect, you also need to give ``rate``\n effect with desired sampling rate, because internally, ``speed`` effects only alter sampling\n rate and leave samples untouched.\n\n Args:\n path (path-like object or file-like object):\n effects (List[List[str]]): List of effects.\n normalize (bool, optional):\n When ``True``, this function always return ``float32``, and sample values are\n normalized to ``[-1.0, 1.0]``.\n If input file is integer WAV, giving ``False`` will change the resulting Tensor type to\n integer type. This argument has no effect for formats other\n than integer WAV type.\n channels_first (bool, optional): When True, the returned Tensor has dimension `[channel, time]`.\n Otherwise, the returned Tensor's dimension is `[time, channel]`.\n format (str or None, optional):\n Override the format detection with the given format.\n Providing the argument might help when libsox can not infer the format\n from header or extension,\n\n Returns:\n (Tensor, int): Resulting Tensor and sample rate.\n If ``normalize=True``, the resulting Tensor is always ``float32`` type.\n If ``normalize=False`` and the input audio file is of integer WAV file, then the\n resulting Tensor has corresponding integer type. (Note 24 bit integer type is not supported)\n If ``channels_first=True``, the resulting Tensor has dimension `[channel, time]`,\n otherwise `[time, channel]`.\n\n Example - Basic usage\n >>>\n >>> # Defines the effects to apply\n >>> effects = [\n ... ['gain', '-n'], # normalises to 0dB\n ... ['pitch', '5'], # 5 cent pitch shift\n ... ['rate', '8000'], # resample to 8000 Hz\n ... ]\n >>>\n >>> # Apply effects and load data with channels_first=True\n >>> waveform, sample_rate = apply_effects_file(\"data.wav\", effects, channels_first=True)\n >>>\n >>> # Check the result\n >>> waveform.shape\n paddle.Size([2, 8000])\n >>> waveform\n tensor([[ 5.1151e-03, 1.8073e-02, 2.2188e-02, ..., 1.0431e-07,\n -1.4761e-07, 1.8114e-07],\n [-2.6924e-03, 2.1860e-03, 1.0650e-02, ..., 6.4122e-07,\n -5.6159e-07, 4.8103e-07]])\n >>> sample_rate\n 8000\n\n Example - Apply random speed perturbation to dataset\n >>>\n >>> # Load data from file, apply random speed perturbation\n >>> class RandomPerturbationFile(paddle.utils.data.Dataset):\n ... \\\"\\\"\\\"Given flist, apply random speed perturbation\n ...\n ... Suppose all the input files are at least one second long.\n ... \\\"\\\"\\\"\n ... def __init__(self, flist: List[str], sample_rate: int):\n ... super().__init__()\n ... self.flist = flist\n ... self.sample_rate = sample_rate\n ...\n ... def __getitem__(self, index):\n ... speed = 0.5 + 1.5 * random.randn()\n ... effects = [\n ... ['gain', '-n', '-10'], # apply 10 db attenuation\n ... ['remix', '-'], # merge all the channels\n ... ['speed', f'{speed:.5f}'], # duration is now 0.5 ~ 2.0 seconds.\n ... ['rate', f'{self.sample_rate}'],\n ... ['pad', '0', '1.5'], # add 1.5 seconds silence at the end\n ... ['trim', '0', '2'], # get the first 2 seconds\n ... ]\n ... waveform, _ = paddleaudio.sox_effects.apply_effects_file(\n ... self.flist[index], effects)\n ... return waveform\n ...\n ... def __len__(self):\n ... return len(self.flist)\n ...\n >>> dataset = RandomPerturbationFile(file_list, sample_rate=8000)\n >>> loader = paddle.utils.data.DataLoader(dataset, batch_size=32)\n >>> for batch in loader:\n >>> pass\n \"\"\"\n if hasattr(path, \"read\"):\n ret = paddleaudio._paddleaudio.apply_effects_fileobj(\n path, effects, normalize, channels_first, format)\n if ret is None:\n raise RuntimeError(\"Failed to load audio from {}\".format(path))\n return (paddle.to_tensor(ret[0]), ret[1])\n path = os.fspath(path)\n ret = paddleaudio._paddleaudio.sox_effects_apply_effects_file(\n path, effects, normalize, channels_first, format)\n if ret is not None:\n return (paddle.to_tensor(ret[0]), ret[1])\n raise RuntimeError(\"Failed to load audio from {}\".format(path))\n"
},
{
"path": "audio/paddleaudio/src/CMakeLists.txt",
"content": "if (MSVC)\n set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)\nendif()\n\nif(APPLE)\nset(CMAKE_SHARED_LIBRARY_SUFFIX \".so\")\nendif(APPLE)\n\n################################################################################\n# libpaddleaudio\n################################################################################\nset(\n LIBPADDLEAUDIO_SOURCES\n utils.cpp\n )\n\nset(\n LIBPADDLEAUDIO_INCLUDE_DIRS\n ${PROJECT_SOURCE_DIR}\n )\n\nset(\n LIBPADDLEAUDIO_LINK_LIBRARIES\n )\n\nset(\n LIBPADDLEAUDIO_COMPILE_DEFINITIONS)\n\n#------------------------------------------------------------------------------#\n# START OF CUSTOMIZATION LOGICS\n#------------------------------------------------------------------------------#\n\nif(BUILD_SOX)\n list(\n APPEND\n LIBPADDLEAUDIO_LINK_LIBRARIES\n libsox\n )\n list(\n APPEND\n LIBPADDLEAUDIO_SOURCES\n )\n list(\n APPEND\n LIBPADDLEAUDIO_COMPILE_DEFINITIONS\n INCLUDE_SOX\n )\nendif()\n\n\nif(BUILD_KALDI)\n list(\n APPEND\n LIBPADDLEAUDIO_LINK_LIBRARIES\n kaldi-native-fbank-core\n )\n list(\n APPEND\n LIBPADDLEAUDIO_COMPILE_DEFINITIONS\n INCLUDE_KALDI\n COMPILE_WITHOUT_OPENFST\n )\nendif()\n\n#------------------------------------------------------------------------------#\n# END OF CUSTOMIZATION LOGICS\n#------------------------------------------------------------------------------#\n\nfunction (define_library name source include_dirs link_libraries compile_defs)\n add_library(${name} SHARED ${source})\n target_include_directories(${name} PRIVATE ${include_dirs})\n target_link_libraries(${name} ${link_libraries})\n target_compile_definitions(${name} PRIVATE ${compile_defs})\n set_target_properties(${name} PROPERTIES PREFIX \"\")\n if (MSVC)\n set_target_properties(${name} PROPERTIES SUFFIX \".pyd\")\n endif(MSVC)\n\n install(\n TARGETS ${name}\n LIBRARY DESTINATION lib\n RUNTIME DESTINATION lib # For Windows\n )\nendfunction()\n\n\ndefine_library(\n libpaddleaudio\n \"${LIBPADDLEAUDIO_SOURCES}\"\n \"${LIBPADDLEAUDIO_INCLUDE_DIRS}\"\n \"${LIBPADDLEAUDIO_LINK_LIBRARIES}\"\n \"${LIBPADDLEAUDIO_COMPILE_DEFINITIONS}\"\n)\n\nif (APPLE)\n set(AUDIO_LIBRARY libpaddleaudio CACHE INTERNAL \"\")\nelse()\n set(AUDIO_LIBRARY -Wl,--no-as-needed libpaddleaudio -Wl,--as-needed CACHE INTERNAL \"\")\nendif()\n\n ################################################################################\n# _paddleaudio.so\n################################################################################\nif (BUILD_PADDLEAUDIO_PYTHON_EXTENSION)\nif (WIN32)\n find_package(Python3 ${PYTHON_VERSION} EXACT COMPONENTS Development)\n set(ADDITIONAL_ITEMS Python3::Python)\nendif()\nfunction(define_extension name sources include_dirs libraries definitions)\n add_library(${name} SHARED ${sources})\n target_compile_definitions(${name} PRIVATE \"${definitions}\")\n target_include_directories(\n ${name} PRIVATE ${PROJECT_SOURCE_DIR} ${Python_INCLUDE_DIR} ${pybind11_INCLUDE_DIR} ${include_dirs})\n target_link_libraries(\n ${name}\n ${libraries}\n ${PYTHON_LIBRARY}\n ${ADDITIONAL_ITEMS}\n )\n set_target_properties(${name} PROPERTIES PREFIX \"\")\n if (MSVC)\n set_target_properties(${name} PROPERTIES SUFFIX \".pyd\")\n endif(MSVC)\n if (APPLE)\n # https://github.com/facebookarchive/caffe2/issues/854#issuecomment-364538485\n # https://github.com/pytorch/pytorch/commit/73f6715f4725a0723d8171d3131e09ac7abf0666\n set_target_properties(${name} PROPERTIES LINK_FLAGS \"-undefined dynamic_lookup\")\n endif()\n install(\n TARGETS ${name}\n LIBRARY DESTINATION .\n RUNTIME DESTINATION . # For Windows\n )\nendfunction()\n\nset(\n EXTENSION_SOURCES\n pybind/pybind.cpp\n )\n#----------------------------------------------------------------------------#\n# START OF CUSTOMIZATION LOGICS\n#----------------------------------------------------------------------------#\nif(BUILD_SOX)\n list(\n APPEND\n EXTENSION_SOURCES\n pybind/sox/effects.cpp\n pybind/sox/effects_chain.cpp\n pybind/sox/io.cpp\n pybind/sox/types.cpp\n pybind/sox/utils.cpp\n )\nendif()\n\nif(BUILD_KALDI)\n list(\n APPEND\n EXTENSION_SOURCES\n pybind/kaldi/kaldi_feature_wrapper.cc\n pybind/kaldi/kaldi_feature.cc\n )\nendif()\n#----------------------------------------------------------------------------#\n# END OF CUSTOMIZATION LOGICS\n#----------------------------------------------------------------------------#\ndefine_extension(\n _paddleaudio\n \"${EXTENSION_SOURCES}\"\n \"\"\n libpaddleaudio\n \"${LIBPADDLEAUDIO_COMPILE_DEFINITIONS}\"\n )\n# if(BUILD_CTC_DECODER)\n# set(\n# DECODER_EXTENSION_SOURCES\n# decoder/bindings/pybind.cpp\n# )\n# define_extension(\n# _paddleaudio_decoder\n# \"${DECODER_EXTENSION_SOURCES}\"\n# \"\"\n# \"libpaddleaudio_decoder\"\n# \"${LIBPADDLEAUDIO_DECODER_DEFINITIONS}\"\n# )\n# endif()\n# if(USE_FFMPEG)\n# set(\n# FFMPEG_EXTENSION_SOURCES\n# ffmpeg/pybind/typedefs.cpp\n# ffmpeg/pybind/pybind.cpp\n# ffmpeg/pybind/stream_reader.cpp\n# )\n# define_extension(\n# _paddleaudio_ffmpeg\n# \"${FFMPEG_EXTENSION_SOURCES}\"\n# \"${FFMPEG_INCLUDE_DIRS}\"\n# \"libpaddleaudio_ffmpeg\"\n# \"${LIBPADDLEAUDIO_DECODER_DEFINITIONS}\"\n# )\n# endif()\nendif()\n"
},
{
"path": "audio/paddleaudio/src/optional/COPYING",
"content": "Creative Commons Legal Code\n\nCC0 1.0 Universal\n\n CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT PROVIDE\n LEGAL SERVICES. DISTRIBUTION OF THIS DOCUMENT DOES NOT CREATE AN\n ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS\n INFORMATION ON AN \"AS-IS\" BASIS. CREATIVE COMMONS MAKES NO WARRANTIES\n REGARDING THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS\n PROVIDED HEREUNDER, AND DISCLAIMS LIABILITY FOR DAMAGES RESULTING FROM\n THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS PROVIDED\n HEREUNDER.\n\nStatement of Purpose\n\nThe laws of most jurisdictions throughout the world automatically confer\nexclusive Copyright and Related Rights (defined below) upon the creator\nand subsequent owner(s) (each and all, an \"owner\") of an original work of\nauthorship and/or a database (each, a \"Work\").\n\nCertain owners wish to permanently relinquish those rights to a Work for\nthe purpose of contributing to a commons of creative, cultural and\nscientific works (\"Commons\") that the public can reliably and without fear\nof later claims of infringement build upon, modify, incorporate in other\nworks, reuse and redistribute as freely as possible in any form whatsoever\nand for any purposes, including without limitation commercial purposes.\nThese owners may contribute to the Commons to promote the ideal of a free\nculture and the further production of creative, cultural and scientific\nworks, or to gain reputation or greater distribution for their Work in\npart through the use and efforts of others.\n\nFor these and/or other purposes and motivations, and without any\nexpectation of additional consideration or compensation, the person\nassociating CC0 with a Work (the \"Affirmer\"), to the extent that he or she\nis an owner of Copyright and Related Rights in the Work, voluntarily\nelects to apply CC0 to the Work and publicly distribute the Work under its\nterms, with knowledge of his or her Copyright and Related Rights in the\nWork and the meaning and intended legal effect of CC0 on those rights.\n\n1. Copyright and Related Rights. A Work made available under CC0 may be\nprotected by copyright and related or neighboring rights (\"Copyright and\nRelated Rights\"). Copyright and Related Rights include, but are not\nlimited to, the following:\n\n i. the right to reproduce, adapt, distribute, perform, display,\n communicate, and translate a Work;\n ii. moral rights retained by the original author(s) and/or performer(s);\niii. publicity and privacy rights pertaining to a person's image or\n likeness depicted in a Work;\n iv. rights protecting against unfair competition in regards to a Work,\n subject to the limitations in paragraph 4(a), below;\n v. rights protecting the extraction, dissemination, use and reuse of data\n in a Work;\n vi. database rights (such as those arising under Directive 96/9/EC of the\n European Parliament and of the Council of 11 March 1996 on the legal\n protection of databases, and under any national implementation\n thereof, including any amended or successor version of such\n directive); and\nvii. other similar, equivalent or corresponding rights throughout the\n world based on applicable law or treaty, and any national\n implementations thereof.\n\n2. Waiver. To the greatest extent permitted by, but not in contravention\nof, applicable law, Affirmer hereby overtly, fully, permanently,\nirrevocably and unconditionally waives, abandons, and surrenders all of\nAffirmer's Copyright and Related Rights and associated claims and causes\nof action, whether now known or unknown (including existing as well as\nfuture claims and causes of action), in the Work (i) in all territories\nworldwide, (ii) for the maximum duration provided by applicable law or\ntreaty (including future time extensions), (iii) in any current or future\nmedium and for any number of copies, and (iv) for any purpose whatsoever,\nincluding without limitation commercial, advertising or promotional\npurposes (the \"Waiver\"). Affirmer makes the Waiver for the benefit of each\nmember of the public at large and to the detriment of Affirmer's heirs and\nsuccessors, fully intending that such Waiver shall not be subject to\nrevocation, rescission, cancellation, termination, or any other legal or\nequitable action to disrupt the quiet enjoyment of the Work by the public\nas contemplated by Affirmer's express Statement of Purpose.\n\n3. Public License Fallback. Should any part of the Waiver for any reason\nbe judged legally invalid or ineffective under applicable law, then the\nWaiver shall be preserved to the maximum extent permitted taking into\naccount Affirmer's express Statement of Purpose. In addition, to the\nextent the Waiver is so judged Affirmer hereby grants to each affected\nperson a royalty-free, non transferable, non sublicensable, non exclusive,\nirrevocable and unconditional license to exercise Affirmer's Copyright and\nRelated Rights in the Work (i) in all territories worldwide, (ii) for the\nmaximum duration provided by applicable law or treaty (including future\ntime extensions), (iii) in any current or future medium and for any number\nof copies, and (iv) for any purpose whatsoever, including without\nlimitation commercial, advertising or promotional purposes (the\n\"License\"). The License shall be deemed effective as of the date CC0 was\napplied by Affirmer to the Work. Should any part of the License for any\nreason be judged legally invalid or ineffective under applicable law, such\npartial invalidity or ineffectiveness shall not invalidate the remainder\nof the License, and in such case Affirmer hereby affirms that he or she\nwill not (i) exercise any of his or her remaining Copyright and Related\nRights in the Work or (ii) assert any associated claims and causes of\naction with respect to the Work, in either case contrary to Affirmer's\nexpress Statement of Purpose.\n\n4. Limitations and Disclaimers.\n\n a. No trademark or patent rights held by Affirmer are waived, abandoned,\n surrendered, licensed or otherwise affected by this document.\n b. Affirmer offers the Work as-is and makes no representations or\n warranties of any kind concerning the Work, express, implied,\n statutory or otherwise, including without limitation warranties of\n title, merchantability, fitness for a particular purpose, non\n infringement, or the absence of latent or other defects, accuracy, or\n the present or absence of errors, whether or not discoverable, all to\n the greatest extent permissible under applicable law.\n c. Affirmer disclaims responsibility for clearing rights of other persons\n that may apply to the Work or any use thereof, including without\n limitation any person's Copyright and Related Rights in the Work.\n Further, Affirmer disclaims responsibility for obtaining any necessary\n consents, permissions or other rights required for any use of the\n Work.\n d. Affirmer understands and acknowledges that Creative Commons is not a\n party to this document and has no duty or obligation with respect to\n this CC0 or use of the Work.\n"
},
{
"path": "audio/paddleaudio/src/optional/optional.hpp",
"content": "\n///\n// optional - An implementation of std::optional with extensions\n// Written in 2017 by Sy Brand (tartanllama@gmail.com, @TartanLlama)\n//\n// Documentation available at https://tl.tartanllama.xyz/\n//\n// To the extent possible under law, the author(s) have dedicated all\n// copyright and related and neighboring rights to this software to the\n// public domain worldwide. This software is distributed without any warranty.\n//\n// You should have received a copy of the CC0 Public Domain Dedication\n// along with this software. If not, see\n// .\n// https://github.com/TartanLlama/optional\n///\n\n#ifndef TL_OPTIONAL_HPP\n#define TL_OPTIONAL_HPP\n\n#define TL_OPTIONAL_VERSION_MAJOR 1\n#define TL_OPTIONAL_VERSION_MINOR 0\n#define TL_OPTIONAL_VERSION_PATCH 0\n\n#include \n#include \n#include \n#include \n#include \n\n#if (defined(_MSC_VER) && _MSC_VER == 1900)\n#define TL_OPTIONAL_MSVC2015\n#endif\n\n#if (defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ <= 9 && \\\n !defined(__clang__))\n#define TL_OPTIONAL_GCC49\n#endif\n\n#if (defined(__GNUC__) && __GNUC__ == 5 && __GNUC_MINOR__ <= 4 && \\\n !defined(__clang__))\n#define TL_OPTIONAL_GCC54\n#endif\n\n#if (defined(__GNUC__) && __GNUC__ == 5 && __GNUC_MINOR__ <= 5 && \\\n !defined(__clang__))\n#define TL_OPTIONAL_GCC55\n#endif\n\n#if (defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ <= 9 && \\\n !defined(__clang__))\n// GCC < 5 doesn't support overloading on const&& for member functions\n#define TL_OPTIONAL_NO_CONSTRR\n\n// GCC < 5 doesn't support some standard C++11 type traits\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_CONSTRUCTIBLE(T) \\\n std::has_trivial_copy_constructor::value\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_ASSIGNABLE(T) \\\n std::has_trivial_copy_assign::value\n\n// This one will be different for GCC 5.7 if it's ever supported\n#define TL_OPTIONAL_IS_TRIVIALLY_DESTRUCTIBLE(T) \\\n std::is_trivially_destructible::value\n\n// GCC 5 < v < 8 has a bug in is_trivially_copy_constructible which breaks\n// std::vector\n// for non-copyable types\n#elif (defined(__GNUC__) && __GNUC__ < 8 && !defined(__clang__))\n#ifndef TL_GCC_LESS_8_TRIVIALLY_COPY_CONSTRUCTIBLE_MUTEX\n#define TL_GCC_LESS_8_TRIVIALLY_COPY_CONSTRUCTIBLE_MUTEX\nnamespace tl {\nnamespace detail {\ntemplate \nstruct is_trivially_copy_constructible\n : std::is_trivially_copy_constructible {};\n#ifdef _GLIBCXX_VECTOR\ntemplate \nstruct is_trivially_copy_constructible>\n : std::is_trivially_copy_constructible {};\n#endif\n}\n}\n#endif\n\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_CONSTRUCTIBLE(T) \\\n tl::detail::is_trivially_copy_constructible::value\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_ASSIGNABLE(T) \\\n std::is_trivially_copy_assignable::value\n#define TL_OPTIONAL_IS_TRIVIALLY_DESTRUCTIBLE(T) \\\n std::is_trivially_destructible::value\n#else\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_CONSTRUCTIBLE(T) \\\n std::is_trivially_copy_constructible::value\n#define TL_OPTIONAL_IS_TRIVIALLY_COPY_ASSIGNABLE(T) \\\n std::is_trivially_copy_assignable::value\n#define TL_OPTIONAL_IS_TRIVIALLY_DESTRUCTIBLE(T) \\\n std::is_trivially_destructible::value\n#endif\n\n#if __cplusplus > 201103L\n#define TL_OPTIONAL_CXX14\n#endif\n\n// constexpr implies const in C++11, not C++14\n#if (__cplusplus == 201103L || defined(TL_OPTIONAL_MSVC2015) || \\\n defined(TL_OPTIONAL_GCC49))\n#define TL_OPTIONAL_11_CONSTEXPR\n#else\n#define TL_OPTIONAL_11_CONSTEXPR constexpr\n#endif\n\nnamespace tl {\n#ifndef TL_MONOSTATE_INPLACE_MUTEX\n#define TL_MONOSTATE_INPLACE_MUTEX\n/// Used to represent an optional with no data; essentially a bool\nclass monostate {};\n\n/// A tag type to tell optional to construct its value in-place\nstruct in_place_t {\n explicit in_place_t() = default;\n};\n/// A tag to tell optional to construct its value in-place\nstatic constexpr in_place_t in_place{};\n#endif\n\ntemplate \nclass optional;\n\nnamespace detail {\n#ifndef TL_TRAITS_MUTEX\n#define TL_TRAITS_MUTEX\n// C++14-style aliases for brevity\ntemplate \nusing remove_const_t = typename std::remove_const::type;\ntemplate \nusing remove_reference_t = typename std::remove_reference::type;\ntemplate \nusing decay_t = typename std::decay::type;\ntemplate \nusing enable_if_t = typename std::enable_if::type;\ntemplate \nusing conditional_t = typename std::conditional::type;\n\n// std::conjunction from C++17\ntemplate \nstruct conjunction : std::true_type {};\ntemplate \nstruct conjunction : B {};\ntemplate \nstruct conjunction\n : std::conditional, B>::type {};\n\n#if defined(_LIBCPP_VERSION) && __cplusplus == 201103L\n#define TL_TRAITS_LIBCXX_MEM_FN_WORKAROUND\n#endif\n\n// In C++11 mode, there's an issue in libc++'s std::mem_fn\n// which results in a hard-error when using it in a noexcept expression\n// in some cases. This is a check to workaround the common failing case.\n#ifdef TL_TRAITS_LIBCXX_MEM_FN_WORKAROUND\ntemplate \nstruct is_pointer_to_non_const_member_func : std::false_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\ntemplate \nstruct is_pointer_to_non_const_member_func\n : std::true_type {};\n\ntemplate \nstruct is_const_or_const_ref : std::false_type {};\ntemplate \nstruct is_const_or_const_ref : std::true_type {};\ntemplate \nstruct is_const_or_const_ref : std::true_type {};\n#endif\n\n// std::invoke from C++17\n// https://stackoverflow.com/questions/38288042/c11-14-invoke-workaround\ntemplate <\n typename Fn,\n typename... Args,\n#ifdef TL_TRAITS_LIBCXX_MEM_FN_WORKAROUND\n typename = enable_if_t::value &&\n is_const_or_const_ref::value)>,\n#endif\n typename = enable_if_t>::value>,\n int = 0>\nconstexpr auto invoke(Fn &&f, Args &&... args) noexcept(\n noexcept(std::mem_fn(f)(std::forward(args)...)))\n -> decltype(std::mem_fn(f)(std::forward(args)...)) {\n return std::mem_fn(f)(std::forward(args)...);\n}\n\ntemplate >::value>>\nconstexpr auto invoke(Fn &&f, Args &&... args) noexcept(\n noexcept(std::forward(f)(std::forward(args)...)))\n -> decltype(std::forward(f)(std::forward(args)...)) {\n return std::forward(f)(std::forward(args)...);\n}\n\n// std::invoke_result from C++17\ntemplate \nstruct invoke_result_impl;\n\ntemplate \nstruct invoke_result_impl<\n F,\n decltype(detail::invoke(std::declval(), std::declval()...), void()),\n Us...> {\n using type =\n decltype(detail::invoke(std::declval(), std::declval()...));\n};\n\ntemplate \nusing invoke_result = invoke_result_impl;\n\ntemplate \nusing invoke_result_t = typename invoke_result::type;\n\n#if defined(_MSC_VER) && _MSC_VER <= 1900\n// TODO make a version which works with MSVC 2015\ntemplate \nstruct is_swappable : std::true_type {};\n\ntemplate \nstruct is_nothrow_swappable : std::true_type {};\n#else\n// https://stackoverflow.com/questions/26744589/what-is-a-proper-way-to-implement-is-swappable-to-test-for-the-swappable-concept\nnamespace swap_adl_tests {\n// if swap ADL finds this then it would call std::swap otherwise (same\n// signature)\nstruct tag {};\n\ntemplate \ntag swap(T &, T &);\ntemplate \ntag swap(T (&a)[N], T (&b)[N]);\n\n// helper functions to test if an unqualified swap is possible, and if it\n// becomes std::swap\ntemplate \nstd::false_type can_swap(...) noexcept(false);\ntemplate (), std::declval()))>\nstd::true_type can_swap(int) noexcept(noexcept(swap(std::declval(),\n std::declval())));\n\ntemplate \nstd::false_type uses_std(...);\ntemplate \nstd::is_same(), std::declval())), tag>\nuses_std(int);\n\ntemplate \nstruct is_std_swap_noexcept\n : std::integral_constant::value &&\n std::is_nothrow_move_assignable::value> {};\n\ntemplate \nstruct is_std_swap_noexcept : is_std_swap_noexcept {};\n\ntemplate \nstruct is_adl_swap_noexcept\n : std::integral_constant(0))> {};\n} // namespace swap_adl_tests\n\ntemplate \nstruct is_swappable\n : std::integral_constant<\n bool,\n decltype(detail::swap_adl_tests::can_swap(0))::value &&\n (!decltype(detail::swap_adl_tests::uses_std(0))::value ||\n (std::is_move_assignable::value &&\n std::is_move_constructible::value))> {};\n\ntemplate \nstruct is_swappable\n : std::integral_constant<\n bool,\n decltype(detail::swap_adl_tests::can_swap(0))::value &&\n (!decltype(\n detail::swap_adl_tests::uses_std(0))::value ||\n is_swappable::value)> {};\n\ntemplate \nstruct is_nothrow_swappable\n : std::integral_constant<\n bool,\n is_swappable::value &&\n ((decltype(detail::swap_adl_tests::uses_std(0))::value\n &&detail::swap_adl_tests::is_std_swap_noexcept::value) ||\n (!decltype(detail::swap_adl_tests::uses_std(0))::value &&\n detail::swap_adl_tests::is_adl_swap_noexcept::value))> {\n};\n#endif\n#endif\n\n// std::void_t from C++17\ntemplate \nstruct voider {\n using type = void;\n};\ntemplate \nusing void_t = typename voider::type;\n\n// Trait for checking if a type is a tl::optional\ntemplate \nstruct is_optional_impl : std::false_type {};\ntemplate \nstruct is_optional_impl> : std::true_type {};\ntemplate \nusing is_optional = is_optional_impl>;\n\n// Change void to tl::monostate\ntemplate \nusing fixup_void = conditional_t::value, monostate, U>;\n\ntemplate >\nusing get_map_return = optional>>;\n\n// Check if invoking F for some Us returns void\ntemplate \nstruct returns_void_impl;\ntemplate \nstruct returns_void_impl>, U...>\n : std::is_void> {};\ntemplate \nusing returns_void = returns_void_impl;\n\ntemplate \nusing enable_if_ret_void = enable_if_t::value>;\n\ntemplate \nusing disable_if_ret_void = enable_if_t::value>;\n\ntemplate \nusing enable_forward_value =\n detail::enable_if_t::value &&\n !std::is_same, in_place_t>::value &&\n !std::is_same, detail::decay_t>::value>;\n\ntemplate \nusing enable_from_other = detail::enable_if_t<\n std::is_constructible::value &&\n !std::is_constructible &>::value &&\n !std::is_constructible &&>::value &&\n !std::is_constructible &>::value &&\n !std::is_constructible &&>::value &&\n !std::is_convertible &, T>::value &&\n !std::is_convertible &&, T>::value &&\n !std::is_convertible &, T>::value &&\n !std::is_convertible &&, T>::value>;\n\ntemplate \nusing enable_assign_forward = detail::enable_if_t<\n !std::is_same, detail::decay_t>::value &&\n !detail::conjunction,\n std::is_same>>::value &&\n std::is_constructible::value && std::is_assignable::value>;\n\ntemplate \nusing enable_assign_from_other = detail::enable_if_t<\n std::is_constructible::value &&\n std::is_assignable::value &&\n !std::is_constructible &>::value &&\n !std::is_constructible &&>::value &&\n !std::is_constructible &>::value &&\n !std::is_constructible &&>::value &&\n !std::is_convertible &, T>::value &&\n !std::is_convertible &&, T>::value &&\n !std::is_convertible &, T>::value &&\n !std::is_convertible &&, T>::value &&\n !std::is_assignable &>::value &&\n !std::is_assignable &&>::value &&\n !std::is_assignable &>::value &&\n !std::is_assignable &&>::value>;\n\n// The storage base manages the actual storage, and correctly propagates\n// trivial destruction from T. This case is for when T is not trivially\n// destructible.\ntemplate ::value>\nstruct optional_storage_base {\n TL_OPTIONAL_11_CONSTEXPR optional_storage_base() noexcept\n : m_dummy(),\n m_has_value(false) {}\n\n template \n TL_OPTIONAL_11_CONSTEXPR optional_storage_base(in_place_t, U &&... u)\n : m_value(std::forward(u)...), m_has_value(true) {}\n\n ~optional_storage_base() {\n if (m_has_value) {\n m_value.~T();\n m_has_value = false;\n }\n }\n\n struct dummy {};\n union {\n dummy m_dummy;\n T m_value;\n };\n\n bool m_has_value;\n};\n\n// This case is for when T is trivially destructible.\ntemplate \nstruct optional_storage_base {\n TL_OPTIONAL_11_CONSTEXPR optional_storage_base() noexcept\n : m_dummy(),\n m_has_value(false) {}\n\n template \n TL_OPTIONAL_11_CONSTEXPR optional_storage_base(in_place_t, U &&... u)\n : m_value(std::forward(u)...), m_has_value(true) {}\n\n // No destructor, so this class is trivially destructible\n\n struct dummy {};\n union {\n dummy m_dummy;\n T m_value;\n };\n\n bool m_has_value = false;\n};\n\n// This base class provides some handy member functions which can be used in\n// further derived classes\ntemplate \nstruct optional_operations_base : optional_storage_base {\n using optional_storage_base::optional_storage_base;\n\n void hard_reset() noexcept {\n get().~T();\n this->m_has_value = false;\n }\n\n template \n void construct(Args &&... args) noexcept {\n new (std::addressof(this->m_value)) T(std::forward(args)...);\n this->m_has_value = true;\n }\n\n template \n void assign(Opt &&rhs) {\n if (this->has_value()) {\n if (rhs.has_value()) {\n this->m_value = std::forward(rhs).get();\n } else {\n this->m_value.~T();\n this->m_has_value = false;\n }\n }\n\n else if (rhs.has_value()) {\n construct(std::forward(rhs).get());\n }\n }\n\n bool has_value() const { return this->m_has_value; }\n\n TL_OPTIONAL_11_CONSTEXPR T &get() & { return this->m_value; }\n TL_OPTIONAL_11_CONSTEXPR const T &get() const & { return this->m_value; }\n TL_OPTIONAL_11_CONSTEXPR T &&get() && { return std::move(this->m_value); }\n#ifndef TL_OPTIONAL_NO_CONSTRR\n constexpr const T &&get() const && { return std::move(this->m_value); }\n#endif\n};\n\n// This class manages conditionally having a trivial copy constructor\n// This specialization is for when T is trivially copy constructible\ntemplate \nstruct optional_copy_base : optional_operations_base {\n using optional_operations_base::optional_operations_base;\n};\n\n// This specialization is for when T is not trivially copy constructible\ntemplate \nstruct optional_copy_base : optional_operations_base {\n using optional_operations_base::optional_operations_base;\n\n optional_copy_base() = default;\n optional_copy_base(const optional_copy_base &rhs)\n : optional_operations_base() {\n if (rhs.has_value()) {\n this->construct(rhs.get());\n } else {\n this->m_has_value = false;\n }\n }\n\n optional_copy_base(optional_copy_base &&rhs) = default;\n optional_copy_base &operator=(const optional_copy_base &rhs) = default;\n optional_copy_base &operator=(optional_copy_base &&rhs) = default;\n};\n\n// This class manages conditionally having a trivial move constructor\n// Unfortunately there's no way to achieve this in GCC < 5 AFAIK, since it\n// doesn't implement an analogue to std::is_trivially_move_constructible. We\n// have to make do with a non-trivial move constructor even if T is trivially\n// move constructible\n#ifndef TL_OPTIONAL_GCC49\ntemplate ::value>\nstruct optional_move_base : optional_copy_base {\n using optional_copy_base::optional_copy_base;\n};\n#else\ntemplate \nstruct optional_move_base;\n#endif\ntemplate \nstruct optional_move_base : optional_copy_base {\n using optional_copy_base::optional_copy_base;\n\n optional_move_base() = default;\n optional_move_base(const optional_move_base &rhs) = default;\n\n optional_move_base(optional_move_base &&rhs) noexcept(\n std::is_nothrow_move_constructible::value) {\n if (rhs.has_value()) {\n this->construct(std::move(rhs.get()));\n } else {\n this->m_has_value = false;\n }\n }\n optional_move_base &operator=(const optional_move_base &rhs) = default;\n optional_move_base &operator=(optional_move_base &&rhs) = default;\n};\n\n// This class manages conditionally having a trivial copy assignment operator\ntemplate \nstruct optional_copy_assign_base : optional_move_base {\n using optional_move_base::optional_move_base;\n};\n\ntemplate \nstruct optional_copy_assign_base : optional_move_base {\n using optional_move_base::optional_move_base;\n\n optional_copy_assign_base() = default;\n optional_copy_assign_base(const optional_copy_assign_base &rhs) = default;\n\n optional_copy_assign_base(optional_copy_assign_base &&rhs) = default;\n optional_copy_assign_base &operator=(const optional_copy_assign_base &rhs) {\n this->assign(rhs);\n return *this;\n }\n optional_copy_assign_base &operator=(optional_copy_assign_base &&rhs) =\n default;\n};\n\n// This class manages conditionally having a trivial move assignment operator\n// Unfortunately there's no way to achieve this in GCC < 5 AFAIK, since it\n// doesn't implement an analogue to std::is_trivially_move_assignable. We have\n// to make do with a non-trivial move assignment operator even if T is trivially\n// move assignable\n#ifndef TL_OPTIONAL_GCC49\ntemplate ::value\n &&std::is_trivially_move_constructible::value\n &&std::is_trivially_move_assignable::value>\nstruct optional_move_assign_base : optional_copy_assign_base {\n using optional_copy_assign_base::optional_copy_assign_base;\n};\n#else\ntemplate \nstruct optional_move_assign_base;\n#endif\n\ntemplate \nstruct optional_move_assign_base : optional_copy_assign_base {\n using optional_copy_assign_base::optional_copy_assign_base;\n\n optional_move_assign_base() = default;\n optional_move_assign_base(const optional_move_assign_base &rhs) = default;\n\n optional_move_assign_base(optional_move_assign_base &&rhs) = default;\n\n optional_move_assign_base &operator=(const optional_move_assign_base &rhs) =\n default;\n\n optional_move_assign_base &\n operator=(optional_move_assign_base &&rhs) noexcept(\n std::is_nothrow_move_constructible::value\n &&std::is_nothrow_move_assignable::value) {\n this->assign(std::move(rhs));\n return *this;\n }\n};\n\n// optional_delete_ctor_base will conditionally delete copy and move\n// constructors depending on whether T is copy/move constructible\ntemplate ::value,\n bool EnableMove = std::is_move_constructible::value>\nstruct optional_delete_ctor_base {\n optional_delete_ctor_base() = default;\n optional_delete_ctor_base(const optional_delete_ctor_base &) = default;\n optional_delete_ctor_base(optional_delete_ctor_base &&) noexcept = default;\n optional_delete_ctor_base &operator=(const optional_delete_ctor_base &) =\n default;\n optional_delete_ctor_base &operator=(\n optional_delete_ctor_base &&) noexcept = default;\n};\n\ntemplate \nstruct optional_delete_ctor_base {\n optional_delete_ctor_base() = default;\n optional_delete_ctor_base(const optional_delete_ctor_base &) = default;\n optional_delete_ctor_base(optional_delete_ctor_base &&) noexcept = delete;\n optional_delete_ctor_base &operator=(const optional_delete_ctor_base &) =\n default;\n optional_delete_ctor_base &operator=(\n optional_delete_ctor_base &&) noexcept = default;\n};\n\ntemplate \nstruct optional_delete_ctor_base {\n optional_delete_ctor_base() = default;\n optional_delete_ctor_base(const optional_delete_ctor_base &) = delete;\n optional_delete_ctor_base(optional_delete_ctor_base &&) noexcept = default;\n optional_delete_ctor_base &operator=(const optional_delete_ctor_base &) =\n default;\n optional_delete_ctor_base &operator=(\n optional_delete_ctor_base &&) noexcept = default;\n};\n\ntemplate \nstruct optional_delete_ctor_base {\n optional_delete_ctor_base() = default;\n optional_delete_ctor_base(const optional_delete_ctor_base &) = delete;\n optional_delete_ctor_base(optional_delete_ctor_base &&) noexcept = delete;\n optional_delete_ctor_base &operator=(const optional_delete_ctor_base &) =\n default;\n optional_delete_ctor_base &operator=(\n optional_delete_ctor_base &&) noexcept = default;\n};\n\n// optional_delete_assign_base will conditionally delete copy and move\n// constructors depending on whether T is copy/move constructible + assignable\ntemplate ::value &&\n std::is_copy_assignable::value),\n bool EnableMove = (std::is_move_constructible::value &&\n std::is_move_assignable::value)>\nstruct optional_delete_assign_base {\n optional_delete_assign_base() = default;\n optional_delete_assign_base(const optional_delete_assign_base &) = default;\n optional_delete_assign_base(optional_delete_assign_base &&) noexcept =\n default;\n optional_delete_assign_base &operator=(\n const optional_delete_assign_base &) = default;\n optional_delete_assign_base &operator=(\n optional_delete_assign_base &&) noexcept = default;\n};\n\ntemplate \nstruct optional_delete_assign_base {\n optional_delete_assign_base() = default;\n optional_delete_assign_base(const optional_delete_assign_base &) = default;\n optional_delete_assign_base(optional_delete_assign_base &&) noexcept =\n default;\n optional_delete_assign_base &operator=(\n const optional_delete_assign_base &) = default;\n optional_delete_assign_base &operator=(\n optional_delete_assign_base &&) noexcept = delete;\n};\n\ntemplate \nstruct optional_delete_assign_base {\n optional_delete_assign_base() = default;\n optional_delete_assign_base(const optional_delete_assign_base &) = default;\n optional_delete_assign_base(optional_delete_assign_base &&) noexcept =\n default;\n optional_delete_assign_base &operator=(\n const optional_delete_assign_base &) = delete;\n optional_delete_assign_base &operator=(\n optional_delete_assign_base &&) noexcept = default;\n};\n\ntemplate \nstruct optional_delete_assign_base {\n optional_delete_assign_base() = default;\n optional_delete_assign_base(const optional_delete_assign_base &) = default;\n optional_delete_assign_base(optional_delete_assign_base &&) noexcept =\n default;\n optional_delete_assign_base &operator=(\n const optional_delete_assign_base &) = delete;\n optional_delete_assign_base &operator=(\n optional_delete_assign_base &&) noexcept = delete;\n};\n\n} // namespace detail\n\n/// A tag type to represent an empty optional\nstruct nullopt_t {\n struct do_not_use {};\n constexpr explicit nullopt_t(do_not_use, do_not_use) noexcept {}\n};\n/// Represents an empty optional\nstatic constexpr nullopt_t nullopt{nullopt_t::do_not_use{},\n nullopt_t::do_not_use{}};\n\nclass bad_optional_access : public std::exception {\n public:\n bad_optional_access() = default;\n const char *what() const noexcept { return \"Optional has no value\"; }\n};\n\n/// An optional object is an object that contains the storage for another\n/// object and manages the lifetime of this contained object, if any. The\n/// contained object may be initialized after the optional object has been\n/// initialized, and may be destroyed before the optional object has been\n/// destroyed. The initialization state of the contained object is tracked by\n/// the optional object.\ntemplate \nclass optional : private detail::optional_move_assign_base,\n private detail::optional_delete_ctor_base,\n private detail::optional_delete_assign_base {\n using base = detail::optional_move_assign_base;\n\n static_assert(!std::is_same::value,\n \"instantiation of optional with in_place_t is ill-formed\");\n static_assert(!std::is_same, nullopt_t>::value,\n \"instantiation of optional with nullopt_t is ill-formed\");\n\n public:\n// The different versions for C++14 and 11 are needed because deduced return\n// types are not SFINAE-safe. This provides better support for things like\n// generic lambdas. C.f.\n// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0826r0.html\n#if defined(TL_OPTIONAL_CXX14) && !defined(TL_OPTIONAL_GCC49) && \\\n !defined(TL_OPTIONAL_GCC54) && !defined(TL_OPTIONAL_GCC55)\n /// Carries out some operation which returns an optional on the stored\n /// object if there is one.\n template ![](\"https://ai-studio-static-online.cdn.bcebos.com/06fd746ab32042f398fb6f33f873e6869e846fe63c214596ae37860fe8103720\")