multi-speaker-tacotron

This is an implementation of our paper to appear at ICASSP 2020:
"Zero-Shot Multi-Speaker Text-To-Speech with State-of-the-art Neural Speaker Embeddings," by Erica Cooper, Cheng-I Lai, Yusuke Yasuda, Fuming Fang, Xin Wang, Nanxin Chen, and Junichi Yamagishi.
https://arxiv.org/abs/1910.10838
Please cite this paper if you use this code.

Audio samples can be found here: https://nii-yamagishilab.github.io/samples-multi-speaker-tacotron/

News:

2021-06-21: Added scripts for creating tfrecords to synthesize new texts using pretrained models. See directory synthesize_new_texts and its README.
2020-08-10: Added example scripts for our new paper accepted to Interspeech 2020, "Can Speaker Augmentation Improve Multi-Speaker End-to-End TTS?" See directory is20 and please also update your copies of tacotron2 and self-attention-tacotron repositories as these contain some necessary changes.

Dependencies:

It is recommended to set up a miniconda environment for using Tacotron. https://repo.anaconda.com

conda create -n taco python=3.6.8
conda activate taco
conda install tensorflow-gpu scipy matplotlib docopt hypothesis pyspark unidecode
conda install -c conda-forge librosa
pip install inflect pysptk

Install this repository

git clone https://github.com/nii-yamagishilab/multi-speaker-tacotron.git external/multi_speaker_tacotron

Install Tacotron dependencies if you don't have them already:

mkdir external
git clone https://github.com/nii-yamagishilab/tacotron2.git external/tacotron2
git clone https://github.com/nii-yamagishilab/self-attention-tacotron.git external/self_attention_tacotron

Note the renaming of hyphens to underscores; this is necessary because “-” is an invalid character in Python.

Next, download project data and models, from the dropbox folder here: https://www.dropbox.com/sh/rq4lebus0n8tmso/AACldbmKDPRN9YiXrRROjtTSa?dl=0

Preprocessed VCTK data: in the data directory
VCTK Tacotron models: in the tacotron-models directory
VCTK Wavenet models: in the wavenet-models directory
Nancy model for parameter initialization: TBA

Training from scratch using the VCTK data only is possible using the script train_from_scratch.sh; this does not require the Nancy pre-trained model.

To use our pre-trained WaveNet models, you will also need our WaveNet implementation which can be found here: https://github.com/nii-yamagishilab/project-CURRENNT-scripts

To obtain embeddings for new samples, you will need the neural speaker embedding code which can be found here: https://github.com/jefflai108/pytorch-kaldi-neural-speaker-embeddings

How to use

See the scripts warmup.sh (warm start training), train_from_scratch.sh (train on VCTK data only), and predictmel.sh (prediction). The scripts assume a SLURM-type computing environment. You will need to change the paths to match your environments and point to your data. Here are the parameters relevant to multi-speaker TTS:

source-data-root and target-data-root: path to your source and target preprocessed data
selected-list-dir: train/eval/test set definitions
batch_size: if you get OOM errors, try reducing the batch size
use_external_speaker_embedding=True: use speaker embeddings that you provide from a file (see the files in the speaker_embeddings directory)
embedding_file: path to the file containing your speaker embeddings
speaker_embedding_dim: dimension should match the dimension in your embedding file
speaker_embedding_projection_out_dim=64: We found experimentally that projecting the speaker embedding to a lower dimension helped to reduce overfitting. You can try different values, but to use our pretrained multi-speaker models you will have to use 64.
speaker_embedding_offset: must match the ID of your first speaker.

The scripts are set up using embedding_file="vctk-x-vector.txt",speaker_embedding_dim='200' which is default x-vectors. Please change it to embedding_file="vctk-lde-3.txt",speaker_embedding_dim='512' to use LDE embeddings from our best system.

Acknowledgments

This work was partially supported by a JST CREST Grant (JPMJCR18A6, VoicePersonae project), Japan, and by MEXT KAKENHI Grants (16H06302, 17H04687, 18H04120, 18H04112, 18KT0051, 19K24372), Japan. The numerical calculations were carried out on the TSUBAME 3.0 supercomputer at the Tokyo Institute of Technology.

Licence

BSD 3-Clause License

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

zhaoyun630 / multi-speaker-tacotron