[
{
"path": ".gitignore",
"content": "__pycache__\n.eggs\n*.egg-info\n*.log\n*.out\n*.wav\n*.h5\n.venv\n\negs/*/exp"
},
{
"path": "LICENSE",
"content": "MIT License\n\nCopyright (c) 2022 Reo YONEYAMA\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": "# Source-Filter HiFi-GAN (SiFi-GAN)\n\nThis repo provides official PyTorch implementation of [SiFi-GAN](https://arxiv.org/abs/2210.15533), a fast and pitch controllable high-fidelity neural vocoder.
\nFor more information, please see our [DEMO](https://chomeyama.github.io/SiFiGAN-Demo/).\n\n## Environment setup\n\n```bash\n$ cd SiFiGAN\n$ pip install -e .\n```\n\nPlease refer to the [Parallel WaveGAN](https://github.com/kan-bayashi/ParallelWaveGAN) repo for more details.\n\n## Folder architecture\n- **egs**:\nThe folder for projects.\n- **egs/namine_ritsu**:\nThe folder of the [Namine Ritsu](https://www.youtube.com/watch?v=pKeo9IE_L1I) project example.\n- **sifigan**:\nThe folder of the source codes.\n\nThe dataset preparation of Namine Ritsu database is based on [NNSVS](https://github.com/nnsvs/nnsvs/).\nPlease refer to it for the procedure and details.\n\n## Run\n\nIn this repo, hyperparameters are managed using [Hydra](https://hydra.cc/docs/intro/).
\nHydra provides an easy way to dynamically create a hierarchical configuration by composition and override it through config files and the command line.\n\n### Dataset preparation\n\nMake dataset and scp files denoting paths to each audio files according to your own dataset (e.g., `egs/namine_ritsu/data/scp/namine_ritsu.scp`).
\nList files denoting paths to the extracted features are automatically created in the next step (e.g., `egs/namine_ritsu/data/scp/namine_ritsu.list`).
\nNote that scp/list files for training/validation/evaluation are needed.\n\n### Preprocessing\n\n```bash\n# Move to the project directory\n$ cd egs/namine_ritsu\n\n# Extract acoustic features (F0, mel-cepstrum, and etc.)\n# You can customize parameters according to sifigan/bin/config/extract_features.yaml\n$ sifigan-extract-features audio=data/scp/namine_ritsu_all.scp\n\n# Compute statistics of training data\n$ sifigan-compute-statistics feats=data/scp/namine_ritsu_train.list stats=data/stats/namine_ritsu_train.joblib\n```\n\n### Training\n\n```bash\n# Train a model customizing the hyperparameters as you like\n$ sifigan-train generator=sifigan discriminator=univnet train=sifigan data=namine_ritsu out_dir=exp/sifigan\n```\n\n### Inference\n\n```bash\n# Decode with several F0 scaling factors\n$ sifigan-decode generator=sifigan data=namine_ritsu out_dir=exp/sifigan checkpoint_steps=400000 f0_factors=[0.5,1.0,2.0]\n```\n\n### Analysis-Synthesis\n\n```bash\n# WORLD analysis + Neural vocoder synthesis\n$ sifigan-anasyn generator=sifigan in_dir=your_own_input_wav_dir out_dir=your_own_output_wav_dir stats=pretrained_sifigan/namine_ritsu_train_no_dev.joblib checkpoint_path=pretrained_sifigan/checkpoint-400000steps.pkl f0_factors=[1.0]\n```\n\n### Pretrained model\n\n~~I provide a pretrained SiFiGAN model [HERE](https://www.dropbox.com/s/akofngycxxz1dg5/pretrained_sifigan.tar.gz?dl=0) which is trained on the Namine Ritsu corpus in the same training manner described in the paper.\nYou can download and place it in your own directory. Then set the appropriate path to the pretrained model and the command should work.~~\n\n\n~~However, since the Namine Ritsu corpus includes a single female Japanese singer, there is a possibility that the model would not work well especially for male singers.\nI am planning to publish another pretrained model trained on larger dataset including many speakers.~~\n\nDue to being trained on the code before bug fixes, I have decided to cancel the release of the model trained on the Namine Ritsu database. Instead, a model trained on the following large-scale dataset is available.\n\nA pretrained model on 24 kHz speech + singing datasets is available [HERE](https://drive.google.com/file/d/1uzqTeumvkPQpfdK_D4U41MDL5-s-hs0l/view?usp=sharing). We used train-clean-100 and train-clean-360 in [LibriTTS-R](https://google.github.io/df-conformer/librittsr/), and [NUS-48E](https://www.smcnus.org/wp-content/uploads/2013/09/05-Pub-NUS-48E.pdf) for training.\nTwo speakers, ADIZ and JLEE in NUS-48E, were excluded from the training data for evaluation. Also, the wav data of NUS-48E were divided into clips of approximately one second each before the feature extraction step.\n\nThe feature preprocessing and training commands are as follows:\n```bash\nsifigan-extract-features audio=data/scp/libritts_r_clean+nus-48e_train_no_dev.scp minf0=60 maxf0=1000\nsifigan-extract-features audio=data/scp/libritts_r_clean+nus-48e_dev.scp minf0=60 maxf0=1000\nsifigan-extract-features audio=data/scp/libritts_r_clean+nus-48e_eval.scp minf0=60 maxf0=1000\n\nsifigan-compute-statistics feats=data/scp/libritts_r_clean+nus-48e_train_no_dev.list stats=data/stats/libritts_r_clean+nus-48e_train_no_dev.joblib\n\nsifigan-train out_dir=test/sifigan generator=sifigan data=libritts_r_clean+nus-48e train=sifigan_1000k\n```\n\n### Monitor training progress\n\n```bash\n$ tensorboard --logdir exp\n```\n\n## Citation\nIf you find the code is helpful, please cite the following article.\n\n```\n@INPROCEEDINGS{10095298,\n author={Yoneyama, Reo and Wu, Yi-Chiao and Toda, Tomoki},\n booktitle={ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},\n title={{Source-Filter HiFi-GAN: Fast and Pitch Controllable High-Fidelity Neural Vocoder}},\n year={2023},\n volume={},\n number={},\n pages={1-5},\n doi={10.1109/ICASSP49357.2023.10095298}\n}\n```\n\n## Authors\n\nDevelopment:\n[Reo Yoneyama](https://chomeyama.github.io/Profile/) @ Nagoya University, Japan
\nE-mail: `yoneyama.reo@g.sp.m.is.nagoya-u.ac.jp`\n\nAdvisors:
\n[Yi-Chiao Wu](https://bigpon.github.io/) @ Meta Reality Labs Research, USA
\nE-mail: `yichiaowu@fb.com`
\n[Tomoki Toda](https://sites.google.com/site/tomokitoda/) @ Nagoya University, Japan