search

ZhangXInFD / SpeechTokenizer

This is the code for the SpeechTokenizer presented in the SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models. Samples are presented on
https://0nutation.github.io/SpeechTokenizer.github.io/
Apache License 2.0
487 stars 41 forks source link

readme

SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models

Introduction

This is the code for the SpeechTokenizer presented in the SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models. SpeechTokenizer is a unified speech tokenizer for speech language models, which adopts the Encoder-Decoder architecture with residual vector quantization (RVQ). Unifying semantic and acoustic tokens, SpeechTokenizer disentangles different aspects of speech information hierarchically across different RVQ layers. Specifically, the code indices that the first quantizer of RVQ outputs can be considered as semantic tokens and the output of the remaining quantizers mainly contain timbre info, which serve as supplements for the information lost by the first quantizer. We provide our models:



Overview


The SpeechTokenizer framework.


Welcome to try our SLMTokBench and we will also open source our USLM!

Qick Link

Release

Samples

Samples are provided on our demo page.

Installation

SpeechTokenizer requires Python>=3.8, and a reasonly recent version of PyTorch. To install SpeechTokenizer, you can run from this repository:

pip install -U speechtokenizer

# or you can clone the repo and install locally
git clone https://github.com/ZhangXInFD/SpeechTokenizer.git
cd SpeechTokenizer
pip install .

Model List

Model Dataset Discription
speechtokenizer_hubert_avg LibriSpeech Adopt average representation across all HuBERT layers as semantic teacher
speechtokenizer_snake LibriSpeech + Common Voice Snake activation, average representation across all HuBERT layers

Usage

load model

from speechtokenizer import SpeechTokenizer

config_path = '/path/config.json'
ckpt_path = '/path/SpeechTokenizer.pt'
model = SpeechTokenizer.load_from_checkpoint(config_path, ckpt_path)
model.eval()

Extracting discrete representations

import torchaudio
import torch

# Load and pre-process speech waveform
wav, sr = torchaudio.load('<SPEECH_FILE_PATH>')

# monophonic checking
if wav.shape(0) > 1:
    wav = wav[:1,:]

if sr != model.sample_rate:
    wav = torchaudio.functional.resample(wav, sr, model.sample_rate)

wav = wav.unsqueeze(0)

# Extract discrete codes from SpeechTokenizer
with torch.no_grad():
    codes = model.encode(wav) # codes: (n_q, B, T)

RVQ_1 = codes[:1, :, :] # Contain content info, can be considered as semantic tokens
RVQ_supplement = codes[1:, :, :] # Contain timbre info, complete info lost by the first quantizer

Decoding discrete representations

# Concatenating semantic tokens (RVQ_1) and supplementary timbre tokens and then decoding
wav = model.decode(torch.cat([RVQ_1, RVQ_supplement], axis=0))

# Decoding from RVQ-i:j tokens from the ith quantizers to the jth quantizers
wav = model.decode(codes[i: (j + 1)], st=i)

Train SpeechTokenizer

In the following section, we describe how to train a SpeechTokenizer model by using our trainer.

Data Preprocess

To train the SpeechTokenizer, the first step is to extract semantic teacher representations from raw audio waveforms. We provide an example of how to extract HuBERT representations in scripts/hubert_rep_extract.sh. We explain the arguments in the following:

Train

You can use SpeechTokenizerTrainer to train a SpeechTokenizer as follows:

from speechtokenizer import SpeechTokenizer, SpeechTokenizerTrainer
from speechtokenizer.discriminators import MultiPeriodDiscriminator, MultiScaleDiscriminator, MultiScaleSTFTDiscriminator
import json

# Load model and trainer config
with open('<CONFIG_FILE_PATH>') as f:
    cfg = json.load(f)

# Initialize SpeechTokenizer
generator = SpeechTokenizer(cfg)

# Initialize the discriminators. You can add any discriminator that is not yet implemented in this repository, as long as the output format remains consistent with the discriminators in `speechtokenizer.discriminators`.
discriminators = {'mpd':MultiPeriodDiscriminator(), 'msd':MultiScaleDiscriminator(), 'mstftd':MultiScaleSTFTDiscriminator(32)}

# Initialize Trainer
trainer = SpeechTokenizerTrainer(generator=generator,
                                discriminators=discriminators,
                                cfg=cfg)

# Start training
trainer.train()

# Continue training from checkpoints
trainer.continue_train()

We provide example training scripts in scripts/train_example.sh. All arguments for SpeechTokenizerTrainer are defined in config/spt_base_cfg.json. Below, we explain some of the important arguments:

Quick Start

If you want to fully follow our experimental setup, simply set semantic_model_path in config/spt_base_cfg.json, and AUDIO_DIR, REP_DIR, EXTS in scripts/hubert_rep_extract.sh, and other optional arguments , then execute the following code:

cd SpeechTokenizer

# Extact semantic representation
bash scripts/hubert_rep_extract.sh

# Train
bash scripts/train_example.sh

Citation

If you use this code or result in your paper, please cite our work as:

@misc{zhang2023speechtokenizer,
      title={SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models}, 
      author={Xin Zhang and Dong Zhang and Shimin Li and Yaqian Zhou and Xipeng Qiu},
      year={2023},
      eprint={2308.16692},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License

The code in this repository is released under the Apache 2.0 license as found in the LICENSE file.