[
{
"path": ".gitignore",
"content": "__pycache__/\nsuno_bark.egg-info/\n"
},
{
"path": "LICENSE",
"content": "MIT License\n\nCopyright (c) Suno, Inc\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n"
},
{
"path": "README.md",
"content": "> Notice: Bark is Suno's open-source text-to-speech+ model. If you are looking for our text-to-music models, please visit us on our [web page](https://suno.ai) and join our community on [Discord](https://suno.ai/discord). \n\n \n# ๐ถ Bark\n\n[](https://suno.ai/discord)\n[](https://twitter.com/suno_ai_)\n\n> ๐ [Examples](https://suno.ai/examples/bark-v0) โข [Suno Studio Waitlist](https://suno-ai.typeform.com/suno-studio) โข [Updates](#-updates) โข [How to Use](#-usage-in-python) โข [Installation](#-installation) โข [FAQ](#-faq)\n\n[//]:
(vertical spaces around image)\n
\n\n![](\"https://user-images.githubusercontent.com/5068315/235310676-a4b3b511-90ec-4edf-8153-7ccf14905d73.png\")
\n
\n
\n\nBark is a transformer-based text-to-audio model created by [Suno](https://suno.ai). Bark can generate highly realistic, multilingual speech as well as other audio - including music, background noise and simple sound effects. The model can also produce nonverbal communications like laughing, sighing and crying. To support the research community, we are providing access to pretrained model checkpoints, which are ready for inference and available for commercial use.\n\n## โ Disclaimer\nBark was developed for research purposes. It is not a conventional text-to-speech model but instead a fully generative text-to-audio model, which can deviate in unexpected ways from provided prompts. Suno does not take responsibility for any output generated. Use at your own risk, and please act responsibly.\n\n## ๐ Quick Index\n* [๐ Updates](#-updates)\n* [๐ป Installation](#-installation)\n* [๐ Usage](#-usage-in-python)\n* [๐ Live Examples](https://suno.ai/examples/bark-v0)\n* [โ FAQ](#-faq)\n\n## ๐ง Demos \n\n[](https://huggingface.co/spaces/suno/bark)\n[](https://replicate.com/suno-ai/bark)\n[](https://colab.research.google.com/drive/1eJfA2XUa-mXwdMy7DoYKVYHI1iTd9Vkt?usp=sharing)\n\n## ๐ Updates\n\n**2023.05.01**\n- ยฉ๏ธ Bark is now licensed under the MIT License, meaning it's now available for commercial use! \n- โก 2x speed-up on GPU. 10x speed-up on CPU. We also added an option for a smaller version of Bark, which offers additional speed-up with the trade-off of slightly lower quality. \n- ๐ [Long-form generation](notebooks/long_form_generation.ipynb), voice consistency enhancements and other examples are now documented in a new [notebooks](./notebooks) section.\n- ๐ฅ We created a [voice prompt library](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c). We hope this resource helps you find useful prompts for your use cases! You can also join us on [Discord](https://suno.ai/discord), where the community actively shares useful prompts in the **#audio-prompts** channel. \n- ๐ฌ Growing community support and access to new features here: \n\n [](https://suno.ai/discord)\n\n- ๐พ You can now use Bark with GPUs that have low VRAM (<4GB).\n\n**2023.04.20**\n- ๐ถ Bark release!\n\n## ๐ Usage in Python\n\n\n ๐ช Basics
\n\n```python\nfrom bark import SAMPLE_RATE, generate_audio, preload_models\nfrom scipy.io.wavfile import write as write_wav\nfrom IPython.display import Audio\n\n# download and load all models\npreload_models()\n\n# generate audio from text\ntext_prompt = \"\"\"\n Hello, my name is Suno. And, uh โ and I like pizza. [laughs] \n But I also have other interests such as playing tic tac toe.\n\"\"\"\naudio_array = generate_audio(text_prompt)\n\n# save audio to disk\nwrite_wav(\"bark_generation.wav\", SAMPLE_RATE, audio_array)\n \n# play text in notebook\nAudio(audio_array, rate=SAMPLE_RATE)\n```\n \n[pizza1.webm](https://user-images.githubusercontent.com/34592747/cfa98e54-721c-4b9c-b962-688e09db684f.webm)\n\n \n\n\n ๐ Foreign Language
\n
\nBark supports various languages out-of-the-box and automatically determines language from input text. When prompted with code-switched text, Bark will attempt to employ the native accent for the respective languages. English quality is best for the time being, and we expect other languages to further improve with scaling. \n
\n
\n\n```python\n\ntext_prompt = \"\"\"\n ์ถ์์ ๋ด๊ฐ ๊ฐ์ฅ ์ข์ํ๋ ๋ช
์ ์ด๋ค. ๋๋ ๋ฉฐ์น ๋์ ํด์์ ์ทจํ๊ณ ์น๊ตฌ ๋ฐ ๊ฐ์กฑ๊ณผ ์๊ฐ์ ๋ณด๋ผ ์ ์์ต๋๋ค.\n\"\"\"\naudio_array = generate_audio(text_prompt)\n```\n[suno_korean.webm](https://user-images.githubusercontent.com/32879321/235313033-dc4477b9-2da0-4b94-9c8b-a8c2d8f5bb5e.webm)\n \n*Note: since Bark recognizes languages automatically from input text, it is possible to use, for example, a german history prompt with english text. This usually leads to english audio with a german accent.*\n```python\ntext_prompt = \"\"\"\n Der Dreiรigjรคhrige Krieg (1618-1648) war ein verheerender Konflikt, der Europa stark geprรคgt hat.\n This is a beginning of the history. If you want to hear more, please continue.\n\"\"\"\naudio_array = generate_audio(text_prompt)\n```\n[suno_german_accent.webm](https://user-images.githubusercontent.com/34592747/3f96ab3e-02ec-49cb-97a6-cf5af0b3524a.webm)\n\n\n \n\n \n\n\n ๐ถ Music
\nBark can generate all types of audio, and, in principle, doesn't see a difference between speech and music. Sometimes Bark chooses to generate text as music, but you can help it out by adding music notes around your lyrics.\n
\n
\n\n```python\ntext_prompt = \"\"\"\n โช In the jungle, the mighty jungle, the lion barks tonight โช\n\"\"\"\naudio_array = generate_audio(text_prompt)\n```\n[lion.webm](https://user-images.githubusercontent.com/5068315/230684766-97f5ea23-ad99-473c-924b-66b6fab24289.webm)\n \n\n\n๐ค Voice Presets
\n \nBark supports 100+ speaker presets across [supported languages](#supported-languages). You can browse the library of supported voice presets [HERE](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c), or in the [code](bark/assets/prompts). The community also often shares presets in [Discord](https://discord.gg/J2B2vsjKuE).\n\n> Bark tries to match the tone, pitch, emotion and prosody of a given preset, but does not currently support custom voice cloning. The model also attempts to preserve music, ambient noise, etc.\n\n```python\ntext_prompt = \"\"\"\n I have a silky smooth voice, and today I will tell you about \n the exercise regimen of the common sloth.\n\"\"\"\naudio_array = generate_audio(text_prompt, history_prompt=\"v2/en_speaker_1\")\n```\n\n[sloth.webm](https://user-images.githubusercontent.com/5068315/230684883-a344c619-a560-4ff5-8b99-b4463a34487b.webm)\n \n\n### ๐ Generating Longer Audio\n \nBy default, `generate_audio` works well with around 13 seconds of spoken text. For an example of how to do long-form generation, see ๐ **[Notebook](notebooks/long_form_generation.ipynb)** ๐\n\n\nClick to toggle example long-form generations (from the example notebook)
\n\n[dialog.webm](https://user-images.githubusercontent.com/2565833/235463539-f57608da-e4cb-4062-8771-148e29512b01.webm)\n\n[longform_advanced.webm](https://user-images.githubusercontent.com/2565833/235463547-1c0d8744-269b-43fe-9630-897ea5731652.webm)\n\n[longform_basic.webm](https://user-images.githubusercontent.com/2565833/235463559-87efe9f8-a2db-4d59-b764-57db83f95270.webm)\n\n \n\n\n## Command line\n```commandline\npython -m bark --text \"Hello, my name is Suno.\" --output_filename \"example.wav\"\n```\n\n## ๐ป Installation\n*โผ๏ธ CAUTION โผ๏ธ Do NOT use `pip install bark`. It installs a different package, which is not managed by Suno.*\n```bash\npip install git+https://github.com/suno-ai/bark.git\n```\n\nor\n\n```bash\ngit clone https://github.com/suno-ai/bark\ncd bark && pip install . \n```\n\n\n## ๐ค Transformers Usage\n\nBark is available in the ๐ค Transformers library from version 4.31.0 onwards, requiring minimal dependencies \nand additional packages. Steps to get started:\n\n1. First install the ๐ค [Transformers library](https://github.com/huggingface/transformers) from main:\n\n```\npip install git+https://github.com/huggingface/transformers.git\n```\n\n2. Run the following Python code to generate speech samples:\n\n```py\nfrom transformers import AutoProcessor, BarkModel\n\nprocessor = AutoProcessor.from_pretrained(\"suno/bark\")\nmodel = BarkModel.from_pretrained(\"suno/bark\")\n\nvoice_preset = \"v2/en_speaker_6\"\n\ninputs = processor(\"Hello, my dog is cute\", voice_preset=voice_preset)\n\naudio_array = model.generate(**inputs)\naudio_array = audio_array.cpu().numpy().squeeze()\n```\n\n3. Listen to the audio samples either in an ipynb notebook:\n\n```py\nfrom IPython.display import Audio\n\nsample_rate = model.generation_config.sample_rate\nAudio(audio_array, rate=sample_rate)\n```\n\nOr save them as a `.wav` file using a third-party library, e.g. `scipy`:\n\n```py\nimport scipy\n\nsample_rate = model.generation_config.sample_rate\nscipy.io.wavfile.write(\"bark_out.wav\", rate=sample_rate, data=audio_array)\n```\n\nFor more details on using the Bark model for inference using the ๐ค Transformers library, refer to the \n[Bark docs](https://huggingface.co/docs/transformers/main/en/model_doc/bark) or the hands-on \n[Google Colab](https://colab.research.google.com/drive/1dWWkZzvu7L9Bunq9zvD-W02RFUXoW-Pd?usp=sharing).\n\n\n## ๐ ๏ธ Hardware and Inference Speed\n\nBark has been tested and works on both CPU and GPU (`pytorch 2.0+`, CUDA 11.7 and CUDA 12.0).\n\nOn enterprise GPUs and PyTorch nightly, Bark can generate audio in roughly real-time. On older GPUs, default colab, or CPU, inference time might be significantly slower. For older GPUs or CPU you might want to consider using smaller models. Details can be found in out tutorial sections here.\n\nThe full version of Bark requires around 12GB of VRAM to hold everything on GPU at the same time. \nTo use a smaller version of the models, which should fit into 8GB VRAM, set the environment flag `SUNO_USE_SMALL_MODELS=True`.\n\nIf you don't have hardware available or if you want to play with bigger versions of our models, you can also sign up for early access to our model playground [here](https://suno-ai.typeform.com/suno-studio).\n\n## โ๏ธ Details\n\nBark is fully generative text-to-audio model devolved for research and demo purposes. It follows a GPT style architecture similar to [AudioLM](https://arxiv.org/abs/2209.03143) and [Vall-E](https://arxiv.org/abs/2301.02111) and a quantized Audio representation from [EnCodec](https://github.com/facebookresearch/encodec). It is not a conventional TTS model, but instead a fully generative text-to-audio model capable of deviating in unexpected ways from any given script. Different to previous approaches, the input text prompt is converted directly to audio without the intermediate use of phonemes. It can therefore generalize to arbitrary instructions beyond speech such as music lyrics, sound effects or other non-speech sounds.\n\nBelow is a list of some known non-speech sounds, but we are finding more every day. Please let us know if you find patterns that work particularly well on [Discord](https://suno.ai/discord)!\n\n- `[laughter]`\n- `[laughs]`\n- `[sighs]`\n- `[music]`\n- `[gasps]`\n- `[clears throat]`\n- `โ` or `...` for hesitations\n- `โช` for song lyrics\n- CAPITALIZATION for emphasis of a word\n- `[MAN]` and `[WOMAN]` to bias Bark toward male and female speakers, respectively\n\n### Supported Languages\n\n| Language | Status |\n| --- | :---: |\n| English (en) | โ
|\n| German (de) | โ
|\n| Spanish (es) | โ
|\n| French (fr) | โ
|\n| Hindi (hi) | โ
|\n| Italian (it) | โ
|\n| Japanese (ja) | โ
|\n| Korean (ko) | โ
|\n| Polish (pl) | โ
|\n| Portuguese (pt) | โ
|\n| Russian (ru) | โ
|\n| Turkish (tr) | โ
|\n| Chinese, simplified (zh) | โ
|\n\nRequests for future language support [here](https://github.com/suno-ai/bark/discussions/111) or in the **#forums** channel on [Discord](https://suno.ai/discord). \n\n## ๐ Appreciation\n\n- [nanoGPT](https://github.com/karpathy/nanoGPT) for a dead-simple and blazing fast implementation of GPT-style models\n- [EnCodec](https://github.com/facebookresearch/encodec) for a state-of-the-art implementation of a fantastic audio codec\n- [AudioLM](https://github.com/lucidrains/audiolm-pytorch) for related training and inference code\n- [Vall-E](https://arxiv.org/abs/2301.02111), [AudioLM](https://arxiv.org/abs/2209.03143) and many other ground-breaking papers that enabled the development of Bark\n\n## ยฉ License\n\nBark is licensed under the MIT License. \n\n## ๐ฑย Community\n\n- [Twitter](https://twitter.com/suno_ai_)\n- [Discord](https://suno.ai/discord)\n\n## ๐งย Suno Studio (Early Access)\n\nWeโre developing a playground for our models, including Bark. \n\nIf you are interested, you can sign up for early access [here](https://suno-ai.typeform.com/suno-studio).\n\n## โ FAQ\n\n#### How do I specify where models are downloaded and cached?\n* Bark uses Hugging Face to download and store models. You can see find more info [here](https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables#hfhome). \n\n\n#### Bark's generations sometimes differ from my prompts. What's happening?\n* Bark is a GPT-style model. As such, it may take some creative liberties in its generations, resulting in higher-variance model outputs than traditional text-to-speech approaches.\n\n#### What voices are supported by Bark? \n* Bark supports 100+ speaker presets across [supported languages](#supported-languages). You can browse the library of speaker presets [here](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c). The community also shares presets in [Discord](https://suno.ai/discord). Bark also supports generating unique random voices that fit the input text. Bark does not currently support custom voice cloning.\n\n#### Why is the output limited to ~13-14 seconds?\n* Bark is a GPT-style model, and its architecture/context window is optimized to output generations with roughly this length.\n\n#### How much VRAM do I need?\n* The full version of Bark requires around 12Gb of memory to hold everything on GPU at the same time. However, even smaller cards down to ~2Gb work with some additional settings. Simply add the following code snippet before your generation: \n\n```python\nimport os\nos.environ[\"SUNO_OFFLOAD_CPU\"] = \"True\"\nos.environ[\"SUNO_USE_SMALL_MODELS\"] = \"True\"\n```\n\n#### My generated audio sounds like a 1980s phone call. What's happening?\n* Bark generates audio from scratch. It is not meant to create only high-fidelity, studio-quality speech. Rather, outputs could be anything from perfect speech to multiple people arguing at a baseball game recorded with bad microphones.\n"
},
{
"path": "bark/__init__.py",
"content": "from .api import generate_audio, text_to_semantic, semantic_to_waveform, save_as_prompt\nfrom .generation import SAMPLE_RATE, preload_models\n"
},
{
"path": "bark/__main__.py",
"content": "from .cli import cli\n\ncli()\n"
},
{
"path": "bark/api.py",
"content": "from typing import Dict, Optional, Union\n\nimport numpy as np\n\nfrom .generation import codec_decode, generate_coarse, generate_fine, generate_text_semantic\n\n\ndef text_to_semantic(\n text: str,\n history_prompt: Optional[Union[Dict, str]] = None,\n temp: float = 0.7,\n silent: bool = False,\n):\n \"\"\"Generate semantic array from text.\n\n Args:\n text: text to be turned into audio\n history_prompt: history choice for audio cloning\n temp: generation temperature (1.0 more diverse, 0.0 more conservative)\n silent: disable progress bar\n\n Returns:\n numpy semantic array to be fed into `semantic_to_waveform`\n \"\"\"\n x_semantic = generate_text_semantic(\n text,\n history_prompt=history_prompt,\n temp=temp,\n silent=silent,\n use_kv_caching=True\n )\n return x_semantic\n\n\ndef semantic_to_waveform(\n semantic_tokens: np.ndarray,\n history_prompt: Optional[Union[Dict, str]] = None,\n temp: float = 0.7,\n silent: bool = False,\n output_full: bool = False,\n):\n \"\"\"Generate audio array from semantic input.\n\n Args:\n semantic_tokens: semantic token output from `text_to_semantic`\n history_prompt: history choice for audio cloning\n temp: generation temperature (1.0 more diverse, 0.0 more conservative)\n silent: disable progress bar\n output_full: return full generation to be used as a history prompt\n\n Returns:\n numpy audio array at sample frequency 24khz\n \"\"\"\n coarse_tokens = generate_coarse(\n semantic_tokens,\n history_prompt=history_prompt,\n temp=temp,\n silent=silent,\n use_kv_caching=True\n )\n fine_tokens = generate_fine(\n coarse_tokens,\n history_prompt=history_prompt,\n temp=0.5,\n )\n audio_arr = codec_decode(fine_tokens)\n if output_full:\n full_generation = {\n \"semantic_prompt\": semantic_tokens,\n \"coarse_prompt\": coarse_tokens,\n \"fine_prompt\": fine_tokens,\n }\n return full_generation, audio_arr\n return audio_arr\n\n\ndef save_as_prompt(filepath, full_generation):\n assert(filepath.endswith(\".npz\"))\n assert(isinstance(full_generation, dict))\n assert(\"semantic_prompt\" in full_generation)\n assert(\"coarse_prompt\" in full_generation)\n assert(\"fine_prompt\" in full_generation)\n np.savez(filepath, **full_generation)\n\n\ndef generate_audio(\n text: str,\n history_prompt: Optional[Union[Dict, str]] = None,\n text_temp: float = 0.7,\n waveform_temp: float = 0.7,\n silent: bool = False,\n output_full: bool = False,\n):\n \"\"\"Generate audio array from input text.\n\n Args:\n text: text to be turned into audio\n history_prompt: history choice for audio cloning\n text_temp: generation temperature (1.0 more diverse, 0.0 more conservative)\n waveform_temp: generation temperature (1.0 more diverse, 0.0 more conservative)\n silent: disable progress bar\n output_full: return full generation to be used as a history prompt\n\n Returns:\n numpy audio array at sample frequency 24khz\n \"\"\"\n semantic_tokens = text_to_semantic(\n text,\n history_prompt=history_prompt,\n temp=text_temp,\n silent=silent,\n )\n out = semantic_to_waveform(\n semantic_tokens,\n history_prompt=history_prompt,\n temp=waveform_temp,\n silent=silent,\n output_full=output_full,\n )\n if output_full:\n full_generation, audio_arr = out\n return full_generation, audio_arr\n else:\n audio_arr = out\n return audio_arr\n"
},
{
"path": "bark/assets/prompts/readme.md",
"content": "# Example Prompts Data\n\n## Version Two\nThe `v2` prompts are better engineered to follow text with a consistent voice.\nTo use them, simply include `v2` in the prompt. For example\n```python\nfrom bark import generate_audio\ntext_prompt = \"madam I'm adam\"\naudio_array = generate_audio(text_prompt, history_prompt=\"v2/en_speaker_1\")\n```\n\n## Prompt Format\nThe provided data is in the .npz format, which is a file format used in Python for storing arrays and data. The data contains three arrays: semantic_prompt, coarse_prompt, and fine_prompt.\n\n```semantic_prompt```\n\nThe semantic_prompt array contains a sequence of token IDs generated by the BERT tokenizer from Hugging Face. These tokens encode the text input and are used as an input to generate the audio output. The shape of this array is (n,), where n is the number of tokens in the input text.\n\n```coarse_prompt```\n\nThe coarse_prompt array is an intermediate output of the text-to-speech pipeline, and contains token IDs generated by the first two codebooks of the EnCodec Codec from Facebook. This step converts the semantic tokens into a different representation that is better suited for the subsequent step. The shape of this array is (2, m), where m is the number of tokens after conversion by the EnCodec Codec.\n\n```fine_prompt```\n\nThe fine_prompt array is a further processed output of the pipeline, and contains 8 codebooks from the EnCodec Codec. These codebooks represent the final stage of tokenization, and the resulting tokens are used to generate the audio output. The shape of this array is (8, p), where p is the number of tokens after further processing by the EnCodec Codec.\n\nOverall, these arrays represent different stages of a text-to-speech pipeline that converts text input into synthesized audio output. The semantic_prompt array represents the input text, while coarse_prompt and fine_prompt represent intermediate and final stages of tokenization, respectively.\n\n\n\n"
},
{
"path": "bark/cli.py",
"content": "import argparse\nfrom typing import Dict, Optional, Union\nimport os\n\nfrom scipy.io.wavfile import write as write_wav\nfrom .api import generate_audio\nfrom .generation import SAMPLE_RATE\n\n\ndef cli():\n \"\"\"Commandline interface.\"\"\"\n parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n parser.add_argument(\"--text\", type=str, help=\"text to be turned into audio\")\n parser.add_argument(\n \"--output_filename\",\n type=str,\n default=\"bark_generation.wav\",\n help=\"output audio file name\",\n )\n parser.add_argument(\"--output_dir\", type=str, default=\".\", help=\"directory to save the outputs\")\n parser.add_argument(\n \"--history_prompt\",\n type=str,\n default=None,\n help=\"history choice for audio cloning, be path to the .npz file.\",\n )\n parser.add_argument(\n \"--text_temp\",\n default=0.7,\n type=float,\n help=\"generation temperature (1.0 more diverse, 0.0 more conservative)\",\n )\n parser.add_argument(\n \"--waveform_temp\",\n default=0.7,\n type=float,\n help=\"generation temperature (1.0 more diverse, 0.0 more conservative)\",\n )\n parser.add_argument(\"--silent\", default=False, type=bool, help=\"disable progress bar\")\n parser.add_argument(\n \"--output_full\",\n default=False,\n type=bool,\n help=\"return full generation to be used as a history prompt\",\n )\n\n args = vars(parser.parse_args())\n input_text: str = args.get(\"text\")\n output_filename: str = args.get(\"output_filename\")\n output_dir: str = args.get(\"output_dir\")\n history_prompt: str = args.get(\"history_prompt\")\n text_temp: float = args.get(\"text_temp\")\n waveform_temp: float = args.get(\"waveform_temp\")\n silent: bool = args.get(\"silent\")\n output_full: bool = args.get(\"output_full\")\n\n try:\n os.makedirs(output_dir, exist_ok=True)\n generated_audio = generate_audio(\n input_text,\n history_prompt=history_prompt,\n text_temp=text_temp,\n waveform_temp=waveform_temp,\n silent=silent,\n output_full=output_full,\n )\n output_file_path = os.path.join(output_dir, output_filename)\n write_wav(output_file_path, SAMPLE_RATE, generated_audio)\n print(f\"Done! Output audio file is saved at: '{output_file_path}'\")\n except Exception as e:\n print(f\"Oops, an error occurred: {e}\")\n"
},
{
"path": "bark/generation.py",
"content": "import contextlib\nimport gc\nimport os\nimport re\n\nfrom encodec import EncodecModel\nimport funcy\nimport logging\nimport numpy as np\nfrom scipy.special import softmax\nimport torch\nimport torch.nn.functional as F\nimport tqdm\nfrom transformers import BertTokenizer\nfrom huggingface_hub import hf_hub_download\n\nfrom .model import GPTConfig, GPT\nfrom .model_fine import FineGPT, FineGPTConfig\n\nif (\n torch.cuda.is_available() and\n hasattr(torch.cuda, \"amp\") and\n hasattr(torch.cuda.amp, \"autocast\") and\n hasattr(torch.cuda, \"is_bf16_supported\") and\n torch.cuda.is_bf16_supported()\n):\n autocast = funcy.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16)\nelse:\n @contextlib.contextmanager\n def autocast():\n yield\n\n\n# hold models in global scope to lazy load\nglobal models\nmodels = {}\n\nglobal models_devices\nmodels_devices = {}\n\n\nCONTEXT_WINDOW_SIZE = 1024\n\nSEMANTIC_RATE_HZ = 49.9\nSEMANTIC_VOCAB_SIZE = 10_000\n\nCODEBOOK_SIZE = 1024\nN_COARSE_CODEBOOKS = 2\nN_FINE_CODEBOOKS = 8\nCOARSE_RATE_HZ = 75\n\nSAMPLE_RATE = 24_000\n\n\nSUPPORTED_LANGS = [\n (\"English\", \"en\"),\n (\"German\", \"de\"),\n (\"Spanish\", \"es\"),\n (\"French\", \"fr\"),\n (\"Hindi\", \"hi\"),\n (\"Italian\", \"it\"),\n (\"Japanese\", \"ja\"),\n (\"Korean\", \"ko\"),\n (\"Polish\", \"pl\"),\n (\"Portuguese\", \"pt\"),\n (\"Russian\", \"ru\"),\n (\"Turkish\", \"tr\"),\n (\"Chinese\", \"zh\"),\n]\n\nALLOWED_PROMPTS = {\"announcer\"}\nfor _, lang in SUPPORTED_LANGS:\n for prefix in (\"\", f\"v2{os.path.sep}\"):\n for n in range(10):\n ALLOWED_PROMPTS.add(f\"{prefix}{lang}_speaker_{n}\")\n\n\nlogger = logging.getLogger(__name__)\n\n\nCUR_PATH = os.path.dirname(os.path.abspath(__file__))\n\n\ndefault_cache_dir = os.path.join(os.path.expanduser(\"~\"), \".cache\")\nCACHE_DIR = os.path.join(os.getenv(\"XDG_CACHE_HOME\", default_cache_dir), \"suno\", \"bark_v0\")\n\n\ndef _cast_bool_env_var(s):\n return s.lower() in ('true', '1', 't')\n\n\nUSE_SMALL_MODELS = _cast_bool_env_var(os.environ.get(\"SUNO_USE_SMALL_MODELS\", \"False\"))\nGLOBAL_ENABLE_MPS = _cast_bool_env_var(os.environ.get(\"SUNO_ENABLE_MPS\", \"False\"))\nOFFLOAD_CPU = _cast_bool_env_var(os.environ.get(\"SUNO_OFFLOAD_CPU\", \"False\"))\n\n\nREMOTE_MODEL_PATHS = {\n \"text_small\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"text.pt\",\n },\n \"coarse_small\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"coarse.pt\",\n },\n \"fine_small\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"fine.pt\",\n },\n \"text\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"text_2.pt\",\n },\n \"coarse\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"coarse_2.pt\",\n },\n \"fine\": {\n \"repo_id\": \"suno/bark\",\n \"file_name\": \"fine_2.pt\",\n },\n}\n\n\nif not hasattr(torch.nn.functional, 'scaled_dot_product_attention') and torch.cuda.is_available():\n logger.warning(\n \"torch version does not support flash attention. You will get faster\" +\n \" inference speed by upgrade torch to newest nightly version.\"\n )\n\n\ndef _grab_best_device(use_gpu=True):\n if torch.cuda.device_count() > 0 and use_gpu:\n device = \"cuda\"\n elif torch.backends.mps.is_available() and use_gpu and GLOBAL_ENABLE_MPS:\n device = \"mps\"\n else:\n device = \"cpu\"\n return device\n\n\ndef _get_ckpt_path(model_type, use_small=False):\n key = model_type\n if use_small or USE_SMALL_MODELS:\n key += \"_small\"\n return os.path.join(CACHE_DIR, REMOTE_MODEL_PATHS[key][\"file_name\"])\n\n\ndef _download(from_hf_path, file_name):\n os.makedirs(CACHE_DIR, exist_ok=True)\n hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR)\n\n\nclass InferenceContext:\n def __init__(self, benchmark=False):\n # we can't expect inputs to be the same length, so disable benchmarking by default\n self._chosen_cudnn_benchmark = benchmark\n self._cudnn_benchmark = None\n\n def __enter__(self):\n self._cudnn_benchmark = torch.backends.cudnn.benchmark\n torch.backends.cudnn.benchmark = self._chosen_cudnn_benchmark\n\n def __exit__(self, exc_type, exc_value, exc_traceback):\n torch.backends.cudnn.benchmark = self._cudnn_benchmark\n\n\nif torch.cuda.is_available():\n torch.backends.cuda.matmul.allow_tf32 = True\n torch.backends.cudnn.allow_tf32 = True\n\n\n@contextlib.contextmanager\ndef _inference_mode():\n with InferenceContext(), torch.inference_mode(), torch.no_grad(), autocast():\n yield\n\n\ndef _clear_cuda_cache():\n if torch.cuda.is_available():\n torch.cuda.empty_cache()\n torch.cuda.synchronize()\n\n\ndef clean_models(model_key=None):\n global models\n model_keys = [model_key] if model_key is not None else list(models.keys())\n for k in model_keys:\n if k in models:\n del models[k]\n _clear_cuda_cache()\n gc.collect()\n\n\ndef _load_model(ckpt_path, device, use_small=False, model_type=\"text\"):\n if model_type == \"text\":\n ConfigClass = GPTConfig\n ModelClass = GPT\n elif model_type == \"coarse\":\n ConfigClass = GPTConfig\n ModelClass = GPT\n elif model_type == \"fine\":\n ConfigClass = FineGPTConfig\n ModelClass = FineGPT\n else:\n raise NotImplementedError()\n model_key = f\"{model_type}_small\" if use_small or USE_SMALL_MODELS else model_type\n model_info = REMOTE_MODEL_PATHS[model_key]\n if not os.path.exists(ckpt_path):\n logger.info(f\"{model_type} model not found, downloading into `{CACHE_DIR}`.\")\n _download(model_info[\"repo_id\"], model_info[\"file_name\"])\n checkpoint = torch.load(ckpt_path, map_location=device)\n # this is a hack\n model_args = checkpoint[\"model_args\"]\n if \"input_vocab_size\" not in model_args:\n model_args[\"input_vocab_size\"] = model_args[\"vocab_size\"]\n model_args[\"output_vocab_size\"] = model_args[\"vocab_size\"]\n del model_args[\"vocab_size\"]\n gptconf = ConfigClass(**checkpoint[\"model_args\"])\n model = ModelClass(gptconf)\n state_dict = checkpoint[\"model\"]\n # fixup checkpoint\n unwanted_prefix = \"_orig_mod.\"\n for k, v in list(state_dict.items()):\n if k.startswith(unwanted_prefix):\n state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)\n extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())\n extra_keys = set([k for k in extra_keys if not k.endswith(\".attn.bias\")])\n missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())\n missing_keys = set([k for k in missing_keys if not k.endswith(\".attn.bias\")])\n if len(extra_keys) != 0:\n raise ValueError(f\"extra keys found: {extra_keys}\")\n if len(missing_keys) != 0:\n raise ValueError(f\"missing keys: {missing_keys}\")\n model.load_state_dict(state_dict, strict=False)\n n_params = model.get_num_params()\n val_loss = checkpoint[\"best_val_loss\"].item()\n logger.info(f\"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss\")\n model.eval()\n model.to(device)\n del checkpoint, state_dict\n _clear_cuda_cache()\n if model_type == \"text\":\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-multilingual-cased\")\n return {\n \"model\": model,\n \"tokenizer\": tokenizer,\n }\n return model\n\n\ndef _load_codec_model(device):\n model = EncodecModel.encodec_model_24khz()\n model.set_target_bandwidth(6.0)\n model.eval()\n model.to(device)\n _clear_cuda_cache()\n return model\n\n\ndef load_model(use_gpu=True, use_small=False, force_reload=False, model_type=\"text\"):\n _load_model_f = funcy.partial(_load_model, model_type=model_type, use_small=use_small)\n if model_type not in (\"text\", \"coarse\", \"fine\"):\n raise NotImplementedError()\n global models\n global models_devices\n device = _grab_best_device(use_gpu=use_gpu)\n model_key = f\"{model_type}\"\n if OFFLOAD_CPU:\n models_devices[model_key] = device\n device = \"cpu\"\n if model_key not in models or force_reload:\n ckpt_path = _get_ckpt_path(model_type, use_small=use_small)\n clean_models(model_key=model_key)\n model = _load_model_f(ckpt_path, device)\n models[model_key] = model\n if model_type == \"text\":\n models[model_key][\"model\"].to(device)\n else:\n models[model_key].to(device)\n return models[model_key]\n\n\ndef load_codec_model(use_gpu=True, force_reload=False):\n global models\n global models_devices\n device = _grab_best_device(use_gpu=use_gpu)\n if device == \"mps\":\n # encodec doesn't support mps\n device = \"cpu\"\n model_key = \"codec\"\n if OFFLOAD_CPU:\n models_devices[model_key] = device\n device = \"cpu\"\n if model_key not in models or force_reload:\n clean_models(model_key=model_key)\n model = _load_codec_model(device)\n models[model_key] = model\n models[model_key].to(device)\n return models[model_key]\n\n\ndef preload_models(\n text_use_gpu=True,\n text_use_small=False,\n coarse_use_gpu=True,\n coarse_use_small=False,\n fine_use_gpu=True,\n fine_use_small=False,\n codec_use_gpu=True,\n force_reload=False,\n):\n \"\"\"Load all the necessary models for the pipeline.\"\"\"\n if _grab_best_device() == \"cpu\" and (\n text_use_gpu or coarse_use_gpu or fine_use_gpu or codec_use_gpu\n ):\n logger.warning(\"No GPU being used. Careful, inference might be very slow!\")\n _ = load_model(\n model_type=\"text\", use_gpu=text_use_gpu, use_small=text_use_small, force_reload=force_reload\n )\n _ = load_model(\n model_type=\"coarse\",\n use_gpu=coarse_use_gpu,\n use_small=coarse_use_small,\n force_reload=force_reload,\n )\n _ = load_model(\n model_type=\"fine\", use_gpu=fine_use_gpu, use_small=fine_use_small, force_reload=force_reload\n )\n _ = load_codec_model(use_gpu=codec_use_gpu, force_reload=force_reload)\n\n\n####\n# Generation Functionality\n####\n\n\ndef _tokenize(tokenizer, text):\n return tokenizer.encode(text, add_special_tokens=False)\n\n\ndef _detokenize(tokenizer, enc_text):\n return tokenizer.decode(enc_text)\n\n\ndef _normalize_whitespace(text):\n return re.sub(r\"\\s+\", \" \", text).strip()\n\n\nTEXT_ENCODING_OFFSET = 10_048\nSEMANTIC_PAD_TOKEN = 10_000\nTEXT_PAD_TOKEN = 129_595\nSEMANTIC_INFER_TOKEN = 129_599\n\n\ndef _load_history_prompt(history_prompt_input):\n if isinstance(history_prompt_input, str) and history_prompt_input.endswith(\".npz\"):\n history_prompt = np.load(history_prompt_input)\n elif isinstance(history_prompt_input, str):\n # make sure this works on non-ubuntu\n history_prompt_input = os.path.join(*history_prompt_input.split(\"/\"))\n if history_prompt_input not in ALLOWED_PROMPTS:\n raise ValueError(\"history prompt not found\")\n history_prompt = np.load(\n os.path.join(CUR_PATH, \"assets\", \"prompts\", f\"{history_prompt_input}.npz\")\n )\n elif isinstance(history_prompt_input, dict):\n assert(\"semantic_prompt\" in history_prompt_input)\n assert(\"coarse_prompt\" in history_prompt_input)\n assert(\"fine_prompt\" in history_prompt_input)\n history_prompt = history_prompt_input\n else:\n raise ValueError(\"history prompt format unrecognized\")\n return history_prompt\n\n\ndef generate_text_semantic(\n text,\n history_prompt=None,\n temp=0.7,\n top_k=None,\n top_p=None,\n silent=False,\n min_eos_p=0.2,\n max_gen_duration_s=None,\n allow_early_stop=True,\n use_kv_caching=False,\n):\n \"\"\"Generate semantic tokens from text.\"\"\"\n assert isinstance(text, str)\n text = _normalize_whitespace(text)\n assert len(text.strip()) > 0\n if history_prompt is not None:\n history_prompt = _load_history_prompt(history_prompt)\n semantic_history = history_prompt[\"semantic_prompt\"]\n assert (\n isinstance(semantic_history, np.ndarray)\n and len(semantic_history.shape) == 1\n and len(semantic_history) > 0\n and semantic_history.min() >= 0\n and semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1\n )\n else:\n semantic_history = None\n # load models if not yet exist\n global models\n global models_devices\n if \"text\" not in models:\n preload_models()\n model_container = models[\"text\"]\n model = model_container[\"model\"]\n tokenizer = model_container[\"tokenizer\"]\n encoded_text = np.array(_tokenize(tokenizer, text)) + TEXT_ENCODING_OFFSET\n if OFFLOAD_CPU:\n model.to(models_devices[\"text\"])\n device = next(model.parameters()).device\n if len(encoded_text) > 256:\n p = round((len(encoded_text) - 256) / len(encoded_text) * 100, 1)\n logger.warning(f\"warning, text too long, lopping of last {p}%\")\n encoded_text = encoded_text[:256]\n encoded_text = np.pad(\n encoded_text,\n (0, 256 - len(encoded_text)),\n constant_values=TEXT_PAD_TOKEN,\n mode=\"constant\",\n )\n if semantic_history is not None:\n semantic_history = semantic_history.astype(np.int64)\n # lop off if history is too long, pad if needed\n semantic_history = semantic_history[-256:]\n semantic_history = np.pad(\n semantic_history,\n (0, 256 - len(semantic_history)),\n constant_values=SEMANTIC_PAD_TOKEN,\n mode=\"constant\",\n )\n else:\n semantic_history = np.array([SEMANTIC_PAD_TOKEN] * 256)\n x = torch.from_numpy(\n np.hstack([\n encoded_text, semantic_history, np.array([SEMANTIC_INFER_TOKEN])\n ]).astype(np.int64)\n )[None]\n assert x.shape[1] == 256 + 256 + 1\n with _inference_mode():\n x = x.to(device)\n n_tot_steps = 768\n # custom tqdm updates since we don't know when eos will occur\n pbar = tqdm.tqdm(disable=silent, total=n_tot_steps)\n pbar_state = 0\n tot_generated_duration_s = 0\n kv_cache = None\n for n in range(n_tot_steps):\n if use_kv_caching and kv_cache is not None:\n x_input = x[:, [-1]]\n else:\n x_input = x\n logits, kv_cache = model(\n x_input, merge_context=True, use_cache=use_kv_caching, past_kv=kv_cache\n )\n relevant_logits = logits[0, 0, :SEMANTIC_VOCAB_SIZE]\n if allow_early_stop:\n relevant_logits = torch.hstack(\n (relevant_logits, logits[0, 0, [SEMANTIC_PAD_TOKEN]]) # eos\n )\n if top_p is not None:\n # faster to convert to numpy\n original_device = relevant_logits.device\n relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()\n sorted_indices = np.argsort(relevant_logits)[::-1]\n sorted_logits = relevant_logits[sorted_indices]\n cumulative_probs = np.cumsum(softmax(sorted_logits))\n sorted_indices_to_remove = cumulative_probs > top_p\n sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()\n sorted_indices_to_remove[0] = False\n relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf\n relevant_logits = torch.from_numpy(relevant_logits)\n relevant_logits = relevant_logits.to(original_device)\n if top_k is not None:\n v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))\n relevant_logits[relevant_logits < v[-1]] = -float(\"Inf\")\n probs = F.softmax(relevant_logits / temp, dim=-1)\n item_next = torch.multinomial(probs, num_samples=1).to(torch.int32)\n if allow_early_stop and (\n item_next == SEMANTIC_VOCAB_SIZE\n or (min_eos_p is not None and probs[-1] >= min_eos_p)\n ):\n # eos found, so break\n pbar.update(n - pbar_state)\n break\n x = torch.cat((x, item_next[None]), dim=1)\n tot_generated_duration_s += 1 / SEMANTIC_RATE_HZ\n if max_gen_duration_s is not None and tot_generated_duration_s > max_gen_duration_s:\n pbar.update(n - pbar_state)\n break\n if n == n_tot_steps - 1:\n pbar.update(n - pbar_state)\n break\n del logits, relevant_logits, probs, item_next\n\n if n > pbar_state:\n if n > pbar.total:\n pbar.total = n\n pbar.update(n - pbar_state)\n pbar_state = n\n pbar.total = n\n pbar.refresh()\n pbar.close()\n out = x.detach().cpu().numpy().squeeze()[256 + 256 + 1 :]\n if OFFLOAD_CPU:\n model.to(\"cpu\")\n assert all(0 <= out) and all(out < SEMANTIC_VOCAB_SIZE)\n _clear_cuda_cache()\n return out\n\n\ndef _flatten_codebooks(arr, offset_size=CODEBOOK_SIZE):\n assert len(arr.shape) == 2\n arr = arr.copy()\n if offset_size is not None:\n for n in range(1, arr.shape[0]):\n arr[n, :] += offset_size * n\n flat_arr = arr.ravel(\"F\")\n return flat_arr\n\n\nCOARSE_SEMANTIC_PAD_TOKEN = 12_048\nCOARSE_INFER_TOKEN = 12_050\n\n\ndef generate_coarse(\n x_semantic,\n history_prompt=None,\n temp=0.7,\n top_k=None,\n top_p=None,\n silent=False,\n max_coarse_history=630, # min 60 (faster), max 630 (more context)\n sliding_window_len=60,\n use_kv_caching=False,\n):\n \"\"\"Generate coarse audio codes from semantic tokens.\"\"\"\n assert (\n isinstance(x_semantic, np.ndarray)\n and len(x_semantic.shape) == 1\n and len(x_semantic) > 0\n and x_semantic.min() >= 0\n and x_semantic.max() <= SEMANTIC_VOCAB_SIZE - 1\n )\n assert 60 <= max_coarse_history <= 630\n assert max_coarse_history + sliding_window_len <= 1024 - 256\n semantic_to_coarse_ratio = COARSE_RATE_HZ / SEMANTIC_RATE_HZ * N_COARSE_CODEBOOKS\n max_semantic_history = int(np.floor(max_coarse_history / semantic_to_coarse_ratio))\n if history_prompt is not None:\n history_prompt = _load_history_prompt(history_prompt)\n x_semantic_history = history_prompt[\"semantic_prompt\"]\n x_coarse_history = history_prompt[\"coarse_prompt\"]\n assert (\n isinstance(x_semantic_history, np.ndarray)\n and len(x_semantic_history.shape) == 1\n and len(x_semantic_history) > 0\n and x_semantic_history.min() >= 0\n and x_semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1\n and isinstance(x_coarse_history, np.ndarray)\n and len(x_coarse_history.shape) == 2\n and x_coarse_history.shape[0] == N_COARSE_CODEBOOKS\n and x_coarse_history.shape[-1] >= 0\n and x_coarse_history.min() >= 0\n and x_coarse_history.max() <= CODEBOOK_SIZE - 1\n and (\n round(x_coarse_history.shape[-1] / len(x_semantic_history), 1)\n == round(semantic_to_coarse_ratio / N_COARSE_CODEBOOKS, 1)\n )\n )\n x_coarse_history = _flatten_codebooks(x_coarse_history) + SEMANTIC_VOCAB_SIZE\n # trim histories correctly\n n_semantic_hist_provided = np.min(\n [\n max_semantic_history,\n len(x_semantic_history) - len(x_semantic_history) % 2,\n int(np.floor(len(x_coarse_history) / semantic_to_coarse_ratio)),\n ]\n )\n n_coarse_hist_provided = int(round(n_semantic_hist_provided * semantic_to_coarse_ratio))\n x_semantic_history = x_semantic_history[-n_semantic_hist_provided:].astype(np.int32)\n x_coarse_history = x_coarse_history[-n_coarse_hist_provided:].astype(np.int32)\n # TODO: bit of a hack for time alignment (sounds better)\n x_coarse_history = x_coarse_history[:-2]\n else:\n x_semantic_history = np.array([], dtype=np.int32)\n x_coarse_history = np.array([], dtype=np.int32)\n # load models if not yet exist\n global models\n global models_devices\n if \"coarse\" not in models:\n preload_models()\n model = models[\"coarse\"]\n if OFFLOAD_CPU:\n model.to(models_devices[\"coarse\"])\n device = next(model.parameters()).device\n # start loop\n n_steps = int(\n round(\n np.floor(len(x_semantic) * semantic_to_coarse_ratio / N_COARSE_CODEBOOKS)\n * N_COARSE_CODEBOOKS\n )\n )\n assert n_steps > 0 and n_steps % N_COARSE_CODEBOOKS == 0\n x_semantic = np.hstack([x_semantic_history, x_semantic]).astype(np.int32)\n x_coarse = x_coarse_history.astype(np.int32)\n base_semantic_idx = len(x_semantic_history)\n with _inference_mode():\n x_semantic_in = torch.from_numpy(x_semantic)[None].to(device)\n x_coarse_in = torch.from_numpy(x_coarse)[None].to(device)\n n_window_steps = int(np.ceil(n_steps / sliding_window_len))\n n_step = 0\n for _ in tqdm.tqdm(range(n_window_steps), total=n_window_steps, disable=silent):\n semantic_idx = base_semantic_idx + int(round(n_step / semantic_to_coarse_ratio))\n # pad from right side\n x_in = x_semantic_in[:, np.max([0, semantic_idx - max_semantic_history]) :]\n x_in = x_in[:, :256]\n x_in = F.pad(\n x_in,\n (0, 256 - x_in.shape[-1]),\n \"constant\",\n COARSE_SEMANTIC_PAD_TOKEN,\n )\n x_in = torch.hstack(\n [\n x_in,\n torch.tensor([COARSE_INFER_TOKEN])[None].to(device),\n x_coarse_in[:, -max_coarse_history:],\n ]\n )\n kv_cache = None\n for _ in range(sliding_window_len):\n if n_step >= n_steps:\n continue\n is_major_step = n_step % N_COARSE_CODEBOOKS == 0\n\n if use_kv_caching and kv_cache is not None:\n x_input = x_in[:, [-1]]\n else:\n x_input = x_in\n\n logits, kv_cache = model(x_input, use_cache=use_kv_caching, past_kv=kv_cache)\n logit_start_idx = (\n SEMANTIC_VOCAB_SIZE + (1 - int(is_major_step)) * CODEBOOK_SIZE\n )\n logit_end_idx = (\n SEMANTIC_VOCAB_SIZE + (2 - int(is_major_step)) * CODEBOOK_SIZE\n )\n relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]\n if top_p is not None:\n # faster to convert to numpy\n original_device = relevant_logits.device\n relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()\n sorted_indices = np.argsort(relevant_logits)[::-1]\n sorted_logits = relevant_logits[sorted_indices]\n cumulative_probs = np.cumsum(softmax(sorted_logits))\n sorted_indices_to_remove = cumulative_probs > top_p\n sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()\n sorted_indices_to_remove[0] = False\n relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf\n relevant_logits = torch.from_numpy(relevant_logits)\n relevant_logits = relevant_logits.to(original_device)\n if top_k is not None:\n v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))\n relevant_logits[relevant_logits < v[-1]] = -float(\"Inf\")\n probs = F.softmax(relevant_logits / temp, dim=-1)\n item_next = torch.multinomial(probs, num_samples=1).to(torch.int32)\n item_next += logit_start_idx\n x_coarse_in = torch.cat((x_coarse_in, item_next[None]), dim=1)\n x_in = torch.cat((x_in, item_next[None]), dim=1)\n del logits, relevant_logits, probs, item_next\n n_step += 1\n del x_in\n del x_semantic_in\n if OFFLOAD_CPU:\n model.to(\"cpu\")\n gen_coarse_arr = x_coarse_in.detach().cpu().numpy().squeeze()[len(x_coarse_history) :]\n del x_coarse_in\n assert len(gen_coarse_arr) == n_steps\n gen_coarse_audio_arr = gen_coarse_arr.reshape(-1, N_COARSE_CODEBOOKS).T - SEMANTIC_VOCAB_SIZE\n for n in range(1, N_COARSE_CODEBOOKS):\n gen_coarse_audio_arr[n, :] -= n * CODEBOOK_SIZE\n _clear_cuda_cache()\n return gen_coarse_audio_arr\n\n\ndef generate_fine(\n x_coarse_gen,\n history_prompt=None,\n temp=0.5,\n silent=True,\n):\n \"\"\"Generate full audio codes from coarse audio codes.\"\"\"\n assert (\n isinstance(x_coarse_gen, np.ndarray)\n and len(x_coarse_gen.shape) == 2\n and 1 <= x_coarse_gen.shape[0] <= N_FINE_CODEBOOKS - 1\n and x_coarse_gen.shape[1] > 0\n and x_coarse_gen.min() >= 0\n and x_coarse_gen.max() <= CODEBOOK_SIZE - 1\n )\n if history_prompt is not None:\n history_prompt = _load_history_prompt(history_prompt)\n x_fine_history = history_prompt[\"fine_prompt\"]\n assert (\n isinstance(x_fine_history, np.ndarray)\n and len(x_fine_history.shape) == 2\n and x_fine_history.shape[0] == N_FINE_CODEBOOKS\n and x_fine_history.shape[1] >= 0\n and x_fine_history.min() >= 0\n and x_fine_history.max() <= CODEBOOK_SIZE - 1\n )\n else:\n x_fine_history = None\n n_coarse = x_coarse_gen.shape[0]\n # load models if not yet exist\n global models\n global models_devices\n if \"fine\" not in models:\n preload_models()\n model = models[\"fine\"]\n if OFFLOAD_CPU:\n model.to(models_devices[\"fine\"])\n device = next(model.parameters()).device\n # make input arr\n in_arr = np.vstack(\n [\n x_coarse_gen,\n np.zeros((N_FINE_CODEBOOKS - n_coarse, x_coarse_gen.shape[1]))\n + CODEBOOK_SIZE, # padding\n ]\n ).astype(np.int32)\n # prepend history if available (max 512)\n if x_fine_history is not None:\n x_fine_history = x_fine_history.astype(np.int32)\n in_arr = np.hstack(\n [\n x_fine_history[:, -512:].astype(np.int32),\n in_arr,\n ]\n )\n n_history = x_fine_history[:, -512:].shape[1]\n else:\n n_history = 0\n n_remove_from_end = 0\n # need to pad if too short (since non-causal model)\n if in_arr.shape[1] < 1024:\n n_remove_from_end = 1024 - in_arr.shape[1]\n in_arr = np.hstack(\n [\n in_arr,\n np.zeros((N_FINE_CODEBOOKS, n_remove_from_end), dtype=np.int32) + CODEBOOK_SIZE,\n ]\n )\n # we can be lazy about fractional loop and just keep overwriting codebooks\n n_loops = np.max([0, int(np.ceil((x_coarse_gen.shape[1] - (1024 - n_history)) / 512))]) + 1\n with _inference_mode():\n in_arr = torch.tensor(in_arr.T).to(device)\n for n in tqdm.tqdm(range(n_loops), disable=silent):\n start_idx = np.min([n * 512, in_arr.shape[0] - 1024])\n start_fill_idx = np.min([n_history + n * 512, in_arr.shape[0] - 512])\n rel_start_fill_idx = start_fill_idx - start_idx\n in_buffer = in_arr[start_idx : start_idx + 1024, :][None]\n for nn in range(n_coarse, N_FINE_CODEBOOKS):\n logits = model(nn, in_buffer)\n if temp is None:\n relevant_logits = logits[0, rel_start_fill_idx:, :CODEBOOK_SIZE]\n codebook_preds = torch.argmax(relevant_logits, -1)\n else:\n relevant_logits = logits[0, :, :CODEBOOK_SIZE] / temp\n probs = F.softmax(relevant_logits, dim=-1)\n codebook_preds = torch.multinomial(\n probs[rel_start_fill_idx:1024], num_samples=1\n ).reshape(-1)\n codebook_preds = codebook_preds.to(torch.int32)\n in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds\n del logits, codebook_preds\n # transfer over info into model_in and convert to numpy\n for nn in range(n_coarse, N_FINE_CODEBOOKS):\n in_arr[\n start_fill_idx : start_fill_idx + (1024 - rel_start_fill_idx), nn\n ] = in_buffer[0, rel_start_fill_idx:, nn]\n del in_buffer\n gen_fine_arr = in_arr.detach().cpu().numpy().squeeze().T\n del in_arr\n if OFFLOAD_CPU:\n model.to(\"cpu\")\n gen_fine_arr = gen_fine_arr[:, n_history:]\n if n_remove_from_end > 0:\n gen_fine_arr = gen_fine_arr[:, :-n_remove_from_end]\n assert gen_fine_arr.shape[-1] == x_coarse_gen.shape[-1]\n _clear_cuda_cache()\n return gen_fine_arr\n\n\ndef codec_decode(fine_tokens):\n \"\"\"Turn quantized audio codes into audio array using encodec.\"\"\"\n # load models if not yet exist\n global models\n global models_devices\n if \"codec\" not in models:\n preload_models()\n model = models[\"codec\"]\n if OFFLOAD_CPU:\n model.to(models_devices[\"codec\"])\n device = next(model.parameters()).device\n arr = torch.from_numpy(fine_tokens)[None]\n arr = arr.to(device)\n arr = arr.transpose(0, 1)\n emb = model.quantizer.decode(arr)\n out = model.decoder(emb)\n audio_arr = out.detach().cpu().numpy().squeeze()\n del arr, emb, out\n if OFFLOAD_CPU:\n model.to(\"cpu\")\n return audio_arr\n"
},
{
"path": "bark/model.py",
"content": "\"\"\"\nMuch of this code is adapted from Andrej Karpathy's NanoGPT\n(https://github.com/karpathy/nanoGPT)\n\"\"\"\nimport math\nfrom dataclasses import dataclass\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import functional as F\n\nclass LayerNorm(nn.Module):\n \"\"\" LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False \"\"\"\n\n def __init__(self, ndim, bias):\n super().__init__()\n self.weight = nn.Parameter(torch.ones(ndim))\n self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None\n\n def forward(self, input):\n return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)\n\nclass CausalSelfAttention(nn.Module):\n\n def __init__(self, config):\n super().__init__()\n assert config.n_embd % config.n_head == 0\n # key, query, value projections for all heads, but in a batch\n self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)\n # output projection\n self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)\n # regularization\n self.attn_dropout = nn.Dropout(config.dropout)\n self.resid_dropout = nn.Dropout(config.dropout)\n self.n_head = config.n_head\n self.n_embd = config.n_embd\n self.dropout = config.dropout\n # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary\n self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')\n if not self.flash:\n # print(\"WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0\")\n # causal mask to ensure that attention is only applied to the left in the input sequence\n self.register_buffer(\"bias\", torch.tril(torch.ones(config.block_size, config.block_size))\n .view(1, 1, config.block_size, config.block_size))\n\n def forward(self, x, past_kv=None, use_cache=False):\n B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)\n\n # calculate query, key, values for all heads in batch and move head forward to be the batch dim\n q, k ,v = self.c_attn(x).split(self.n_embd, dim=2)\n k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n\n if past_kv is not None:\n past_key = past_kv[0]\n past_value = past_kv[1]\n k = torch.cat((past_key, k), dim=-2)\n v = torch.cat((past_value, v), dim=-2)\n\n FULL_T = k.shape[-2]\n\n if use_cache is True:\n present = (k, v)\n else:\n present = None\n\n # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)\n if self.flash:\n # efficient attention using Flash Attention CUDA kernels\n if past_kv is not None:\n # When `past_kv` is provided, we're doing incremental decoding and `q.shape[2] == 1`: q only contains\n # the query for the last token. scaled_dot_product_attention interprets this as the first token in the\n # sequence, so if is_causal=True it will mask out all attention from it. This is not what we want, so \n # to work around this we set is_causal=False.\n is_causal = False\n else:\n is_causal = True\n\n y = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=is_causal)\n else:\n # manual implementation of attention\n att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))\n att = att.masked_fill(self.bias[:,:,FULL_T-T:FULL_T,:FULL_T] == 0, float('-inf'))\n att = F.softmax(att, dim=-1)\n att = self.attn_dropout(att)\n y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side\n\n # output projection\n y = self.resid_dropout(self.c_proj(y))\n return (y, present)\n\nclass MLP(nn.Module):\n\n def __init__(self, config):\n super().__init__()\n self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)\n self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)\n self.dropout = nn.Dropout(config.dropout)\n self.gelu = nn.GELU()\n\n def forward(self, x):\n x = self.c_fc(x)\n x = self.gelu(x)\n x = self.c_proj(x)\n x = self.dropout(x)\n return x\n\nclass Block(nn.Module):\n\n def __init__(self, config, layer_idx):\n super().__init__()\n self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)\n self.attn = CausalSelfAttention(config)\n self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)\n self.mlp = MLP(config)\n self.layer_idx = layer_idx\n\n def forward(self, x, past_kv=None, use_cache=False):\n attn_output, prev_kvs = self.attn(self.ln_1(x), past_kv=past_kv, use_cache=use_cache)\n x = x + attn_output\n x = x + self.mlp(self.ln_2(x))\n return (x, prev_kvs)\n\n@dataclass\nclass GPTConfig:\n block_size: int = 1024\n input_vocab_size: int = 10_048\n output_vocab_size: int = 10_048\n n_layer: int = 12\n n_head: int = 12\n n_embd: int = 768\n dropout: float = 0.0\n bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster\n\nclass GPT(nn.Module):\n\n def __init__(self, config):\n super().__init__()\n assert config.input_vocab_size is not None\n assert config.output_vocab_size is not None\n assert config.block_size is not None\n self.config = config\n\n self.transformer = nn.ModuleDict(dict(\n wte = nn.Embedding(config.input_vocab_size, config.n_embd),\n wpe = nn.Embedding(config.block_size, config.n_embd),\n drop = nn.Dropout(config.dropout),\n h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),\n ln_f = LayerNorm(config.n_embd, bias=config.bias),\n ))\n self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)\n\n def get_num_params(self, non_embedding=True):\n \"\"\"\n Return the number of parameters in the model.\n For non-embedding count (default), the position embeddings get subtracted.\n The token embeddings would too, except due to the parameter sharing these\n params are actually used as weights in the final layer, so we include them.\n \"\"\"\n n_params = sum(p.numel() for p in self.parameters())\n if non_embedding:\n n_params -= self.transformer.wte.weight.numel()\n n_params -= self.transformer.wpe.weight.numel()\n return n_params\n\n def forward(self, idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False):\n device = idx.device\n b, t = idx.size()\n if past_kv is not None:\n assert t == 1\n tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)\n else:\n if merge_context:\n assert(idx.shape[1] >= 256+256+1)\n t = idx.shape[1] - 256\n else:\n assert t <= self.config.block_size, f\"Cannot forward sequence of length {t}, block size is only {self.config.block_size}\"\n\n # forward the GPT model itself\n if merge_context:\n tok_emb = torch.cat([\n self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),\n self.transformer.wte(idx[:,256+256:])\n ], dim=1)\n else:\n tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)\n\n if past_kv is None:\n past_length = 0\n past_kv = tuple([None] * len(self.transformer.h))\n else:\n past_length = past_kv[0][0].size(-2)\n\n if position_ids is None:\n position_ids = torch.arange(past_length, t + past_length, dtype=torch.long, device=device)\n position_ids = position_ids.unsqueeze(0) # shape (1, t)\n assert position_ids.shape == (1, t)\n\n pos_emb = self.transformer.wpe(position_ids) # position embeddings of shape (1, t, n_embd)\n\n x = self.transformer.drop(tok_emb + pos_emb)\n\n new_kv = () if use_cache else None\n\n for i, (block, past_layer_kv) in enumerate(zip(self.transformer.h, past_kv)):\n x, kv = block(x, past_kv=past_layer_kv, use_cache=use_cache)\n\n if use_cache:\n new_kv = new_kv + (kv,)\n\n x = self.transformer.ln_f(x)\n\n # inference-time mini-optimization: only forward the lm_head on the very last position\n logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim\n\n return (logits, new_kv)\n"
},
{
"path": "bark/model_fine.py",
"content": "\"\"\"\nMuch of this code is adapted from Andrej Karpathy's NanoGPT\n(https://github.com/karpathy/nanoGPT)\n\"\"\"\nfrom dataclasses import dataclass\nimport math\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import functional as F\n\nfrom .model import GPT, GPTConfig, MLP\n\n\nclass NonCausalSelfAttention(nn.Module):\n def __init__(self, config):\n super().__init__()\n assert config.n_embd % config.n_head == 0\n # key, query, value projections for all heads, but in a batch\n self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)\n # output projection\n self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)\n # regularization\n self.attn_dropout = nn.Dropout(config.dropout)\n self.resid_dropout = nn.Dropout(config.dropout)\n self.n_head = config.n_head\n self.n_embd = config.n_embd\n self.dropout = config.dropout\n # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0\n self.flash = (\n hasattr(torch.nn.functional, \"scaled_dot_product_attention\")\n )\n\n def forward(self, x):\n B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)\n\n # calculate query, key, values for all heads in batch and move head forward to be the batch dim\n q, k, v = self.c_attn(x).split(self.n_embd, dim=2)\n k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)\n\n # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)\n if self.flash:\n # efficient attention using Flash Attention CUDA kernels\n y = torch.nn.functional.scaled_dot_product_attention(\n q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=False\n )\n else:\n # manual implementation of attention\n att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))\n att = F.softmax(att, dim=-1)\n att = self.attn_dropout(att)\n y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n y = (\n y.transpose(1, 2).contiguous().view(B, T, C)\n ) # re-assemble all head outputs side by side\n\n # output projection\n y = self.resid_dropout(self.c_proj(y))\n return y\n\n\nclass FineBlock(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.ln_1 = nn.LayerNorm(config.n_embd)\n self.attn = NonCausalSelfAttention(config)\n self.ln_2 = nn.LayerNorm(config.n_embd)\n self.mlp = MLP(config)\n\n def forward(self, x):\n x = x + self.attn(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass FineGPT(GPT):\n def __init__(self, config):\n super().__init__(config)\n del self.lm_head\n self.config = config\n self.n_codes_total = config.n_codes_total\n self.transformer = nn.ModuleDict(\n dict(\n wtes=nn.ModuleList(\n [\n nn.Embedding(config.input_vocab_size, config.n_embd)\n for _ in range(config.n_codes_total)\n ]\n ),\n wpe=nn.Embedding(config.block_size, config.n_embd),\n drop=nn.Dropout(config.dropout),\n h=nn.ModuleList([FineBlock(config) for _ in range(config.n_layer)]),\n ln_f=nn.LayerNorm(config.n_embd),\n )\n )\n self.lm_heads = nn.ModuleList(\n [\n nn.Linear(config.n_embd, config.output_vocab_size, bias=False)\n for _ in range(config.n_codes_given, self.n_codes_total)\n ]\n )\n for i in range(self.n_codes_total - config.n_codes_given):\n self.transformer.wtes[i + 1].weight = self.lm_heads[i].weight\n\n def forward(self, pred_idx, idx):\n device = idx.device\n b, t, codes = idx.size()\n assert (\n t <= self.config.block_size\n ), f\"Cannot forward sequence of length {t}, block size is only {self.config.block_size}\"\n assert pred_idx > 0, \"cannot predict 0th codebook\"\n assert codes == self.n_codes_total, (b, t, codes)\n pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0) # shape (1, t)\n\n # forward the GPT model itself\n tok_embs = [\n wte(idx[:, :, i]).unsqueeze(-1) for i, wte in enumerate(self.transformer.wtes)\n ] # token embeddings of shape (b, t, n_embd)\n tok_emb = torch.cat(tok_embs, dim=-1)\n pos_emb = self.transformer.wpe(pos) # position embeddings of shape (1, t, n_embd)\n x = tok_emb[:, :, :, : pred_idx + 1].sum(dim=-1)\n x = self.transformer.drop(x + pos_emb)\n for block in self.transformer.h:\n x = block(x)\n x = self.transformer.ln_f(x)\n logits = self.lm_heads[pred_idx - self.config.n_codes_given](x)\n return logits\n\n def get_num_params(self, non_embedding=True):\n \"\"\"\n Return the number of parameters in the model.\n For non-embedding count (default), the position embeddings get subtracted.\n The token embeddings would too, except due to the parameter sharing these\n params are actually used as weights in the final layer, so we include them.\n \"\"\"\n n_params = sum(p.numel() for p in self.parameters())\n if non_embedding:\n for wte in self.transformer.wtes:\n n_params -= wte.weight.numel()\n n_params -= self.transformer.wpe.weight.numel()\n return n_params\n\n\n@dataclass\nclass FineGPTConfig(GPTConfig):\n n_codes_total: int = 8\n n_codes_given: int = 1\n"
},
{
"path": "model-card.md",
"content": "# Model Card: Bark\n\nThis is the official codebase for running the text to audio model, from Suno.ai.\n\nThe following is additional information about the models released here. \n\n## Model Details\n\nBark is a series of three transformer models that turn text into audio.\n### Text to semantic tokens\n - Input: text, tokenized with [BERT tokenizer from Hugging Face](https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer)\n - Output: semantic tokens that encode the audio to be generated\n\n### Semantic to coarse tokens\n - Input: semantic tokens\n - Output: tokens from the first two codebooks of the [EnCodec Codec](https://github.com/facebookresearch/encodec) from facebook\n\n### Coarse to fine tokens\n - Input: the first two codebooks from EnCodec\n - Output: 8 codebooks from EnCodec\n\n### Architecture\n| Model | Parameters | Attention | Output Vocab size | \n|:-------------------------:|:----------:|------------|:-----------------:|\n| Text to semantic tokens | 80 M | Causal | 10,000 |\n| Semantic to coarse tokens | 80 M | Causal | 2x 1,024 |\n| Coarse to fine tokens | 80 M | Non-causal | 6x 1,024 |\n\n\n### Release date\nApril 2023\n\n## Broader Implications\nWe anticipate that this model's text to audio capabilities can be used to improve accessbility tools in a variety of languages. \nStraightforward improvements will allow models to run faster than realtime, rendering them useful for applications such as virtual assistants. \n \nWhile we hope that this release will enable users to express their creativity and build applications that are a force\nfor good, we acknowledge that any text to audio model has the potential for dual use. While it is not straightforward\nto voice clone known people with Bark, they can still be used for nefarious purposes. To further reduce the chances of unintended use of Bark, \nwe also release a simple classifier to detect Bark-generated audio with high accuracy (see notebooks section of the main repository). \n"
},
{
"path": "notebooks/fake_classifier.ipynb",
"content": "{\n \"cells\": [\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"e330c1de\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import torchaudio\\n\",\n \"from transformers import HubertModel\\n\",\n \"from sklearn.metrics import PrecisionRecallDisplay\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"2ac3dd88\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# use hubert from HF for feature embedding\\n\",\n \"model = HubertModel.from_pretrained(\\\"facebook/hubert-base-ls960\\\")\\n\",\n \"arr, sr = torchaudio.load(\\\"my_audio.wav\\\")\\n\",\n \"if sr != 16_000:\\n\",\n \" arr = torchaudio.functional.resample(arr, sr, 16_000)\\n\",\n \"# use intermediate layer\\n\",\n \"hidden_state = model(arr[None], output_hidden_states=True).hidden_states[6]\\n\",\n \"# take mean over time\\n\",\n \"feats = hidden_state.detach().cpu().numpy().squeeze().mean(0)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"03a602e0\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# load sk-learn classifier from here: https://dl.suno-models.io/bark/models/v0/classifier.pkl\\n\",\n \"with open(\\\"classifier.pkl\\\", \\\"rb\\\") as f:\\n\",\n \" clf = pickle.load(f)\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"8e423794\",\n \"metadata\": {},\n \"source\": [\n \"### Precision-recall curve on test set\"\n ]\n },\n {\n \"attachments\": {\n \"image.png\": {\n \"image/png\": \"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\"\n }\n },\n \"cell_type\": \"markdown\",\n \"id\": \"e1486424\",\n \"metadata\": {},\n \"source\": [\n \"\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"668856bf\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"c87326bd\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"decdbf09\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3 (ipykernel)\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.8.15\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 5\n}\n"
},
{
"path": "notebooks/long_form_generation.ipynb",
"content": "{\n \"cells\": [\n {\n \"cell_type\": \"code\",\n \"execution_count\": 1,\n \"id\": \"39ea4bed\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import os\\n\",\n \"\\n\",\n \"os.environ[\\\"CUDA_VISIBLE_DEVICES\\\"] = \\\"0\\\"\\n\",\n \"\\n\",\n \"\\n\",\n \"from IPython.display import Audio\\n\",\n \"import nltk # we'll use this to split into sentences\\n\",\n \"import numpy as np\\n\",\n \"\\n\",\n \"from bark.generation import (\\n\",\n \" generate_text_semantic,\\n\",\n \" preload_models,\\n\",\n \")\\n\",\n \"from bark.api import semantic_to_waveform\\n\",\n \"from bark import generate_audio, SAMPLE_RATE\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 29,\n \"id\": \"776964b6\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"preload_models()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"1d03f4d2\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"74a025a4\",\n \"metadata\": {},\n \"source\": [\n \"# Simple Long-Form Generation\\n\",\n \"We split longer text into sentences using `nltk` and generate the sentences one by one.\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 33,\n \"id\": \"57b06e2a\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"script = \\\"\\\"\\\"\\n\",\n \"Hey, have you heard about this new text-to-audio model called \\\"Bark\\\"? \\n\",\n \"Apparently, it's the most realistic and natural-sounding text-to-audio model \\n\",\n \"out there right now. People are saying it sounds just like a real person speaking. \\n\",\n \"I think it uses advanced machine learning algorithms to analyze and understand the \\n\",\n \"nuances of human speech, and then replicates those nuances in its own speech output. \\n\",\n \"It's pretty impressive, and I bet it could be used for things like audiobooks or podcasts. \\n\",\n \"In fact, I heard that some publishers are already starting to use Bark to create audiobooks. \\n\",\n \"It would be like having your own personal voiceover artist. I really think Bark is going to \\n\",\n \"be a game-changer in the world of text-to-audio technology.\\n\",\n \"\\\"\\\"\\\".replace(\\\"\\\\n\\\", \\\" \\\").strip()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 34,\n \"id\": \"f747f804\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"sentences = nltk.sent_tokenize(script)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 35,\n \"id\": \"17400a9b\",\n \"metadata\": {\n \"scrolled\": true\n },\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:02<00:00, 43.03it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 17/17 [00:06<00:00, 2.45it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 22.73it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 33/33 [00:13<00:00, 2.52it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 66.30it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 11/11 [00:04<00:00, 2.46it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 20.99it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 35/35 [00:14<00:00, 2.46it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:03<00:00, 25.63it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 29/29 [00:11<00:00, 2.50it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 23.90it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 30/30 [00:12<00:00, 2.46it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 53.24it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 14/14 [00:05<00:00, 2.51it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 50.63it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 15/15 [00:05<00:00, 2.57it/s]\\n\"\n ]\n }\n ],\n \"source\": [\n \"SPEAKER = \\\"v2/en_speaker_6\\\"\\n\",\n \"silence = np.zeros(int(0.25 * SAMPLE_RATE)) # quarter second of silence\\n\",\n \"\\n\",\n \"pieces = []\\n\",\n \"for sentence in sentences:\\n\",\n \" audio_array = generate_audio(sentence, history_prompt=SPEAKER)\\n\",\n \" pieces += [audio_array, silence.copy()]\\n\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"04cf77f9\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"Audio(np.concatenate(pieces), rate=SAMPLE_RATE)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"ac2d4625\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"6d13249b\",\n \"metadata\": {},\n \"source\": [\n \"# $ \\\\\\\\ $\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"cdfc8bf5\",\n \"metadata\": {},\n \"source\": [\n \"# Advanced Long-Form Generation\\n\",\n \"Somtimes Bark will hallucinate a little extra audio at the end of the prompt.\\n\",\n \"We can solve this issue by lowering the threshold for bark to stop generating text. \\n\",\n \"We use the `min_eos_p` kwarg in `generate_text_semantic`\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 37,\n \"id\": \"62807fd0\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:02<00:00, 38.05it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 18/18 [00:07<00:00, 2.46it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:03<00:00, 32.28it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 21/21 [00:08<00:00, 2.54it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 55.78it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 14/14 [00:05<00:00, 2.57it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:06<00:00, 14.73it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 35/35 [00:14<00:00, 2.47it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:02<00:00, 40.29it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 18/18 [00:07<00:00, 2.56it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:03<00:00, 32.92it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 20/20 [00:08<00:00, 2.47it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 68.87it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 12/12 [00:04<00:00, 2.62it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:02<00:00, 47.64it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 15/15 [00:06<00:00, 2.46it/s]\\n\"\n ]\n }\n ],\n \"source\": [\n \"GEN_TEMP = 0.6\\n\",\n \"SPEAKER = \\\"v2/en_speaker_6\\\"\\n\",\n \"silence = np.zeros(int(0.25 * SAMPLE_RATE)) # quarter second of silence\\n\",\n \"\\n\",\n \"pieces = []\\n\",\n \"for sentence in sentences:\\n\",\n \" semantic_tokens = generate_text_semantic(\\n\",\n \" sentence,\\n\",\n \" history_prompt=SPEAKER,\\n\",\n \" temp=GEN_TEMP,\\n\",\n \" min_eos_p=0.05, # this controls how likely the generation is to end\\n\",\n \" )\\n\",\n \"\\n\",\n \" audio_array = semantic_to_waveform(semantic_tokens, history_prompt=SPEAKER,)\\n\",\n \" pieces += [audio_array, silence.copy()]\\n\",\n \"\\n\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"133fec46\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"Audio(np.concatenate(pieces), rate=SAMPLE_RATE)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"6eee9f5a\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"be8e125e\",\n \"metadata\": {},\n \"source\": [\n \"# $ \\\\\\\\ $\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"03a16c1b\",\n \"metadata\": {},\n \"source\": [\n \"# Make a Long-Form Dialog with Bark\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"06c5eff8\",\n \"metadata\": {},\n \"source\": [\n \"### Step 1: Format a script and speaker lookup\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 14,\n \"id\": \"5238b297\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"['Samantha: Hey, have you heard about this new text-to-audio model called \\\"Bark\\\"?',\\n\",\n \" \\\"John: No, I haven't. What's so special about it?\\\",\\n\",\n \" \\\"Samantha: Well, apparently it's the most realistic and natural-sounding text-to-audio model out there right now. People are saying it sounds just like a real person speaking.\\\",\\n\",\n \" 'John: Wow, that sounds amazing. How does it work?',\\n\",\n \" 'Samantha: I think it uses advanced machine learning algorithms to analyze and understand the nuances of human speech, and then replicates those nuances in its own speech output.',\\n\",\n \" \\\"John: That's pretty impressive. Do you think it could be used for things like audiobooks or podcasts?\\\",\\n\",\n \" 'Samantha: Definitely! In fact, I heard that some publishers are already starting to use Bark to create audiobooks. And I bet it would be great for podcasts too.',\\n\",\n \" 'John: I can imagine. It would be like having your own personal voiceover artist.',\\n\",\n \" 'Samantha: Exactly! I think Bark is going to be a game-changer in the world of text-to-audio technology.']\"\n ]\n },\n \"execution_count\": 14,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"speaker_lookup = {\\\"Samantha\\\": \\\"v2/en_speaker_9\\\", \\\"John\\\": \\\"v2/en_speaker_2\\\"}\\n\",\n \"\\n\",\n \"# Script generated by chat GPT\\n\",\n \"script = \\\"\\\"\\\"\\n\",\n \"Samantha: Hey, have you heard about this new text-to-audio model called \\\"Bark\\\"?\\n\",\n \"\\n\",\n \"John: No, I haven't. What's so special about it?\\n\",\n \"\\n\",\n \"Samantha: Well, apparently it's the most realistic and natural-sounding text-to-audio model out there right now. People are saying it sounds just like a real person speaking.\\n\",\n \"\\n\",\n \"John: Wow, that sounds amazing. How does it work?\\n\",\n \"\\n\",\n \"Samantha: I think it uses advanced machine learning algorithms to analyze and understand the nuances of human speech, and then replicates those nuances in its own speech output.\\n\",\n \"\\n\",\n \"John: That's pretty impressive. Do you think it could be used for things like audiobooks or podcasts?\\n\",\n \"\\n\",\n \"Samantha: Definitely! In fact, I heard that some publishers are already starting to use Bark to create audiobooks. And I bet it would be great for podcasts too.\\n\",\n \"\\n\",\n \"John: I can imagine. It would be like having your own personal voiceover artist.\\n\",\n \"\\n\",\n \"Samantha: Exactly! I think Bark is going to be a game-changer in the world of text-to-audio technology.\\\"\\\"\\\"\\n\",\n \"script = script.strip().split(\\\"\\\\n\\\")\\n\",\n \"script = [s.strip() for s in script if s]\\n\",\n \"script\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"ee547efd\",\n \"metadata\": {},\n \"source\": [\n \"### Step 2: Generate the audio for every speaker turn\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 15,\n \"id\": \"203e5081\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:02<00:00, 34.03it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 22/22 [00:08<00:00, 2.55it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 71.58it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 11/11 [00:04<00:00, 2.65it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 22.75it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 33/33 [00:13<00:00, 2.53it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 70.76it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 11/11 [00:04<00:00, 2.63it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 20.46it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 36/36 [00:14<00:00, 2.47it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 20.18it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 37/37 [00:14<00:00, 2.51it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:04<00:00, 23.04it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 32/32 [00:12<00:00, 2.48it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 54.64it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 14/14 [00:05<00:00, 2.58it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:03<00:00, 31.71it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 24/24 [00:09<00:00, 2.56it/s]\\n\"\n ]\n }\n ],\n \"source\": [\n \"pieces = []\\n\",\n \"silence = np.zeros(int(0.5*SAMPLE_RATE))\\n\",\n \"for line in script:\\n\",\n \" speaker, text = line.split(\\\": \\\")\\n\",\n \" audio_array = generate_audio(text, history_prompt=speaker_lookup[speaker], )\\n\",\n \" pieces += [audio_array, silence.copy()]\"\n ]\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"7c54bada\",\n \"metadata\": {},\n \"source\": [\n \"### Step 3: Concatenate all of the audio and play it\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"27a56842\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"Audio(np.concatenate(pieces), rate=SAMPLE_RATE)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"a1bc5877\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3 (ipykernel)\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.9.16\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 5\n}\n"
},
{
"path": "notebooks/memory_profiling_bark.ipynb",
"content": "{\n \"cells\": [\n {\n \"cell_type\": \"markdown\",\n \"id\": \"90641144\",\n \"metadata\": {},\n \"source\": [\n \"# Bark Memory Profiling\\n\",\n \"Bark has two ways to reduce GPU memory: \\n\",\n \" - Small models: a smaller version of the model. This can be set by using the environment variable `SUNO_USE_SMALL_MODELS`\\n\",\n \" - offloading models to CPU: Holding only one model at a time on the GPU, and shuttling the models to the CPU in between generations. \\n\",\n \"\\n\",\n \"# $ \\\\\\\\ $\\n\",\n \"## First, we'll use the most memory efficient configuration\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 1,\n \"id\": \"39ea4bed\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import os\\n\",\n \"\\n\",\n \"os.environ[\\\"CUDA_VISIBLE_DEVICES\\\"] = \\\"0\\\"\\n\",\n \"os.environ[\\\"SUNO_USE_SMALL_MODELS\\\"] = \\\"1\\\"\\n\",\n \"os.environ[\\\"SUNO_OFFLOAD_CPU\\\"] = \\\"1\\\"\\n\",\n \"\\n\",\n \"from bark.generation import (\\n\",\n \" generate_text_semantic,\\n\",\n \" preload_models,\\n\",\n \")\\n\",\n \"from bark import generate_audio, SAMPLE_RATE\\n\",\n \"\\n\",\n \"import torch\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 2,\n \"id\": \"66b0c006\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:01<00:00, 62.17it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 10/10 [00:03<00:00, 2.74it/s]\\n\"\n ]\n },\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"max memory usage = 2396MB\\n\"\n ]\n }\n ],\n \"source\": [\n \"torch.cuda.reset_peak_memory_stats()\\n\",\n \"preload_models()\\n\",\n \"audio_array = generate_audio(\\\"madam I'm adam\\\", history_prompt=\\\"v2/en_speaker_5\\\")\\n\",\n \"max_utilization = torch.cuda.max_memory_allocated()\\n\",\n \"print(f\\\"max memory usage = {max_utilization / 1024 / 1024:.0f}MB\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"9922dd2d\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"bdbe578e\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"markdown\",\n \"id\": \"213d1b5b\",\n \"metadata\": {},\n \"source\": [\n \"# Memory Profiling:\\n\",\n \"We can profile the memory consumption of 4 scenarios\\n\",\n \" - Small models, offloading to CPU\\n\",\n \" - Large models, offloading to CPU\\n\",\n \" - Small models, not offloading to CPU\\n\",\n \" - Large models, not offloading to CPU\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 1,\n \"id\": \"417d5e9c\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import os\\n\",\n \"\\n\",\n \"from bark.generation import (\\n\",\n \" generate_text_semantic,\\n\",\n \" preload_models,\\n\",\n \" models,\\n\",\n \")\\n\",\n \"import bark.generation\\n\",\n \"\\n\",\n \"from bark.api import semantic_to_waveform\\n\",\n \"from bark import generate_audio, SAMPLE_RATE\\n\",\n \"\\n\",\n \"import torch\\n\",\n \"import time\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 2,\n \"id\": \"cd83b45d\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"Small models True, offloading to CPU: True\\n\",\n \"\\tmax memory usage = 967MB, time 4s\\n\",\n \"\\n\",\n \"Small models False, offloading to CPU: True\\n\",\n \"\\tmax memory usage = 2407MB, time 8s\\n\",\n \"\\n\",\n \"Small models True, offloading to CPU: False\\n\",\n \"\\tmax memory usage = 2970MB, time 3s\\n\",\n \"\\n\",\n \"Small models False, offloading to CPU: False\\n\",\n \"\\tmax memory usage = 7824MB, time 6s\\n\",\n \"\\n\"\n ]\n }\n ],\n \"source\": [\n \"global models\\n\",\n \"\\n\",\n \"for offload_models in (True, False):\\n\",\n \" # this setattr is needed to do on the fly\\n\",\n \" # the easier way to do this is with `os.environ[\\\"SUNO_OFFLOAD_CPU\\\"] = \\\"1\\\"`\\n\",\n \" setattr(bark.generation, \\\"OFFLOAD_CPU\\\", offload_models)\\n\",\n \" for use_small_models in (True, False):\\n\",\n \" models = {}\\n\",\n \" torch.cuda.empty_cache()\\n\",\n \" torch.cuda.reset_peak_memory_stats()\\n\",\n \" preload_models(\\n\",\n \" text_use_small=use_small_models,\\n\",\n \" coarse_use_small=use_small_models,\\n\",\n \" fine_use_small=use_small_models,\\n\",\n \" force_reload=True,\\n\",\n \" )\\n\",\n \" t0 = time.time()\\n\",\n \" audio_array = generate_audio(\\\"madam I'm adam\\\", history_prompt=\\\"v2/en_speaker_5\\\", silent=True)\\n\",\n \" dur = time.time() - t0\\n\",\n \" max_utilization = torch.cuda.max_memory_allocated()\\n\",\n \" print(f\\\"Small models {use_small_models}, offloading to CPU: {offload_models}\\\")\\n\",\n \" print(f\\\"\\\\tmax memory usage = {max_utilization / 1024 / 1024:.0f}MB, time {dur:.0f}s\\\\n\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"bfe5fa06\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3 (ipykernel)\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.9.16\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 5\n}\n"
},
{
"path": "notebooks/use_small_models_on_cpu.ipynb",
"content": "{\n \"cells\": [\n {\n \"cell_type\": \"markdown\",\n \"id\": \"6a682b61\",\n \"metadata\": {},\n \"source\": [\n \"# Benchmarking small models on CPU\\n\",\n \" - We can enable small models with the `SUNO_USE_SMALL_MODELS` environment variable\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 1,\n \"id\": \"9500dd93\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import os\\n\",\n \"\\n\",\n \"os.environ[\\\"CUDA_VISIBLE_DEVICES\\\"] = \\\"\\\"\\n\",\n \"os.environ[\\\"SUNO_USE_SMALL_MODELS\\\"] = \\\"1\\\"\\n\",\n \"\\n\",\n \"from IPython.display import Audio\\n\",\n \"import numpy as np\\n\",\n \"\\n\",\n \"from bark import generate_audio, preload_models, SAMPLE_RATE\\n\",\n \"\\n\",\n \"import time\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 2,\n \"id\": \"4e3454b6\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"No GPU being used. Careful, inference might be very slow!\\n\"\n ]\n },\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"CPU times: user 5.52 s, sys: 2.34 s, total: 7.86 s\\n\",\n \"Wall time: 4.33 s\\n\"\n ]\n }\n ],\n \"source\": [\n \"%%time\\n\",\n \"preload_models()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 3,\n \"id\": \"f6024e5f\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 100/100 [00:10<00:00, 9.89it/s]\\n\",\n \"100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 15/15 [00:43<00:00, 2.90s/it]\\n\"\n ]\n },\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"took 62s to generate 6s of audio\\n\"\n ]\n }\n ],\n \"source\": [\n \"t0 = time.time()\\n\",\n \"text = \\\"In the light of the moon, a little egg lay on a leaf\\\"\\n\",\n \"audio_array = generate_audio(text)\\n\",\n \"generation_duration_s = time.time() - t0\\n\",\n \"audio_duration_s = audio_array.shape[0] / SAMPLE_RATE\\n\",\n \"\\n\",\n \"print(f\\\"took {generation_duration_s:.0f}s to generate {audio_duration_s:.0f}s of audio\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 4,\n \"id\": \"2dcce86c\",\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"10\"\n ]\n },\n \"execution_count\": 4,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"os.cpu_count()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"id\": \"3046eddb\",\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3 (ipykernel)\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.9.16\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 5\n}\n"
},
{
"path": "pyproject.toml",
"content": "[build-system]\nrequires = [\"setuptools\"]\nbuild-backend = \"setuptools.build_meta\"\n\n[project]\nname = \"suno-bark\"\nversion = \"0.0.1a\"\ndescription = \"Bark text to audio model\"\nreadme = \"README.md\"\nrequires-python = \">=3.8\"\nauthors = [\n {name = \"Suno Inc\", email = \"hello@suno.ai\"},\n]\n# Apache 2.0\nlicense = {file = \"LICENSE\"}\n\ndependencies = [\n \"boto3\",\n \"encodec\",\n \"funcy\",\n \"huggingface-hub>=0.14.1\",\n \"numpy\",\n \"scipy\",\n \"tokenizers\",\n \"torch\",\n \"tqdm\",\n \"transformers\",\n]\n\n[project.urls]\nsource = \"https://github.com/suno-ai/bark\"\n\n[project.optional-dependencies]\ndev = [\n \"bandit\",\n \"black\",\n \"codecov\",\n \"flake8\",\n \"hypothesis>=6.14,<7\",\n \"isort>=5.0.0,<6\",\n \"jupyter\",\n \"mypy\",\n \"nbconvert\",\n \"nbformat\",\n \"pydocstyle\",\n \"pylint\",\n \"pytest\",\n \"pytest-cov\",\n]\n\n[tool.setuptools]\npackages = [\"bark\"]\n\n[tool.setuptools.package-data]\nbark = [\"assets/prompts/*.npz\", \"assets/prompts/v2/*.npz\"]\n\n\n[tool.black]\nline-length = 100\n"
},
{
"path": "setup.py",
"content": "from setuptools import setup\n\nsetup()\n"
}
]