vall-e: A Python repository from kevincaicedo

Language : 🇺🇸 | 🇨🇳

An unofficial PyTorch implementation of VALL-E(Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers).

Inference: In-Context Learning via Prompting

model trained with nano config(about 100x smaller than the paper config) can synthesize human-like speech.

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

cd egs/libritts

# VALL-E
# nano config is too small, so the AR-Decoder may not work well.
# re-run to get new(diverse) result.
python3 bin/infer.py \
    --decoder-dim 128 --nhead 4 --num-decoder-layers 4 --model-name valle \
    --text-prompts "Go to her." \
    --audio-prompts ./prompts/61_70970_000007_000001.wav \
    --text "To get up and running quickly just follow the steps below." \
    --output-dir infer/demo_valle_PostNorm_epoch10 \
    --checkpoint exp/valle_nano_v41_PostNorm/epoch-10.pt


# VALL-F is more stable
python3 bin/infer.py \
    --decoder-dim 128 --nhead 4 --num-decoder-layers 4 --model-name vallf \
    --text-prompts "Go to her." \
    --audio-prompts ./prompts/61_70970_000007_000001.wav \
    --text "To get up and running quickly just follow the steps below." \
    --output-dir infer/demo_vallf_PostNorm_epoch10 \
    --checkpoint exp/vallf_nano_v41_PostNorm/epoch-10.pt

VALL-E nano config epoch-20

VALL-F nano config epoch-10

Demo

official demo

Broader impacts

Since VALL-E could synthesize speech that maintains speaker identity, it may carry potential risks in misuse of the model, such as spoofing voice identification or impersonating a specific speaker.

To avoid abuse, Well-trained models and services will not be provided.

Progress

Installation

To get up and running quickly just follow the steps below:

# PyTorch
pip install torch==1.13.1 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

# DeepSpeed
# https://github.com/microsoft/DeepSpeed/issues/2697
sudo apt-get install -y libopenmpi-dev
pip install mpi4py deepspeed==0.7.7

# phonemizer
apt-get install espeak-ng
## OSX: brew install espeak
pip install phonemizer

# lhotse
# https://github.com/lhotse-speech/lhotse/pull/956
# https://github.com/lhotse-speech/lhotse/pull/960
pip uninstall lhotse
pip uninstall lhotse
pip install git+https://github.com/lhotse-speech/lhotse

# k2 icefall
# pip install k2
git clone https://github.com/k2-fsa/k2.git
cd k2
export K2_MAKE_ARGS="-j12"
export K2_CMAKE_ARGS="-DK2_WITH_CUDA=OFF"
python setup.py install
cd -

git clone https://github.com/k2-fsa/icefall
cd icefall
pip install -r requirements.txt
export PYTHONPATH=`pwd`/../icefall:$PYTHONPATH
echo "export PYTHONPATH=`pwd`/../icefall:\$PYTHONPATH" >> ~/.zshrc
echo "export PYTHONPATH=`pwd`/../icefall:\$PYTHONPATH" >> ~/.bashrc
cd -

# valle
git clone https://github.com/lifeiteng/valle.git
cd valle
pip install -e .

Training

egs/libritts/README.md

Troubleshooting

SummaryWriter segmentation fault (core dumped)
- LINE tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
- FIX https://github.com/tensorflow/tensorboard/pull/6135/files

Contributing

Multi-GPU Training
Parallelize bin/tokenizer.py on multi-GPUs
Provide GPU resources (MyEmail: lifeiteng0422@163.com)

Citing

To cite this repository:

@misc{valle,
  author={Feiteng Li},
  title={VALL-E: A neural codec language model},
  year={2023},
  url={http://github.com/lifeiteng/valle}
}

@article{VALL-E,
  title     = {Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers},
  author    = {Chengyi Wang, Sanyuan Chen, Yu Wu,
               Ziqiang Zhang, Long Zhou, Shujie Liu,
               Zhuo Chen, Yanqing Liu, Huaming Wang,
               Jinyu Li, Lei He, Sheng Zhao, Furu Wei},
  year      = {2023},
  eprint    = {2301.02111},
  archivePrefix = {arXiv},
  volume    = {abs/2301.02111},
  url       = {http://arxiv.org/abs/2301.02111},
}

kevincaicedo/vall-e