VALL-E

An unofficial PyTorch implementation of VALL-E, based on the EnCodec tokenizer.

Get Started

Requirements

Since the trainer is based on DeepSpeed, you will need to have a GPU that DeepSpeed has developed and tested against, as well as a CUDA or ROCm compiler pre-installed to install this package.

Install

pip install git+https://github.com/enhuiz/vall-e

Or you may clone by:

git clone --recurse-submodules https://github.com/enhuiz/vall-e.git

Note that the code is only tested under Python 3.10.7.

Training

Put your data into a folder, e.g. data/your_data. Audio files should be named with the suffix .wav and text files with .normalized.txt.
Quantize the data:

python -m vall_e.emb.qnt data/your_data

Generate phonemes based on the text:

python -m vall_e.emb.g2p data/your_data

Customize your configuration by creating config/your_data/ar.yml and config/your_data/nar.yml. Refer to the example configs in config/test and vall_e/config.py for details. You may choose different model presets, check vall_e/vall_e/__init__.py.
Train the AR or NAR model using the following scripts:

python -m vall_e.train yaml=config/your_data/ar_or_nar.yml

You may quit your training any time by just typing quit in your CLI. The latest checkpoint will be automatically saved.

Export trained models:

Both trained models need to be exported to a certain path. To export either of them, run:

python -m vall_e.export zoo/ar_or_nar.pt yaml=config/your_data/ar_or_nar.yml

This will export the latest checkpoint.

Synthesis

python -m vall_e <text> <ref_path> <out_path> --ar-ckpt zoo/ar.pt --nar-ckpt zoo/nar.pt

TODO

AR model for the first quantizer
Audio decoding from tokens
NAR model for the rest quantizers
Trainers for both models
Implement AdaLN for NAR model.
Sample-wise quantization level sampling for NAR training.
Pre-trained checkpoint and demos on LibriTTS
Synthesis CLI

Notice

EnCodec is licensed under CC-BY-NC 4.0. If you use the code to generate audio quantization or perform decoding, it is important to adhere to the terms of their license.

Citations

@article{wang2023neural,
  title={Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers},
  author={Wang, Chengyi and Chen, Sanyuan and Wu, Yu and Zhang, Ziqiang and Zhou, Long and Liu, Shujie and Chen, Zhuo and Liu, Yanqing and Wang, Huaming and Li, Jinyu and others},
  journal={arXiv preprint arXiv:2301.02111},
  year={2023}
}

@article{defossez2022highfi,
  title={High Fidelity Neural Audio Compression},
  author={Défossez, Alexandre and Copet, Jade and Synnaeve, Gabriel and Adi, Yossi},
  journal={arXiv preprint arXiv:2210.13438},
  year={2022}
}

egaebel / vall-e