YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for everyone
In our recent paper we propose the YourTTS model. YourTTS brings the power of a multilingual approach to the task of zero-shot multi-speaker TTS. Our method builds upon the VITS model and adds several novel modifications for zero-shot multi-speaker and multilingual training. We achieved state-of-the-art (SOTA) results in zero-shot multi-speaker TTS and results comparable to SOTA in zero-shot voice conversion on the VCTK dataset. Additionally, our approach achieves promising results in a target language with a single-speaker dataset, opening possibilities for zero-shot multi-speaker TTS and zero-shot voice conversion systems in low-resource languages. Finally, it is possible to fine-tune the YourTTS model with less than 1 minute of speech and achieve state-of-the-art results in voice similarity and with reasonable quality. This is important to allow synthesis for speakers with a very different voice or recording characteristics from those seen during training.
Erratum
In Section 2 of YourTTS paper, we have defined the Speaker Consistency Loss (SCL) function. In addition, we have used this loss function on 4 fine-tuning experiments in Sections 3 and 4 (EXP. 1 + SCL, EXP. 2 + SCL, EXP. 3 + SCL, and EXP. 4 + SCL). However, due to an implementation mistake, the gradient of this loss function was not propagated for the model during the training. It means that the fine-tuning experiments that used this loss are equivalent to training the model for more steps without the Speaker Consistency Loss. This bug was discovered by Tomáš Nekvinda and reported on issue number 2348 of the Coqui TTS repository. This bug was fixed on the pull request number 2364 on the Coqui TTS repository. Currently, it is fixed for Coqui TTS version v0.12.0 or higher. We would like to thank Tomáš Nekvinda for finding the bug and reporting it.
Production version
Come try our latest and greatest fullband english only model https://coqui.ai/
Considering the "target_content_wav.wav" as the reference wave file to convert into the voice of the "target_speaker_wav.wav" speaker.
Results replicability
To insure replicability, we make the audios used to generate the MOS available here. In addition, we provide the MOS for each audio here.
To re-generate our MOS results, follow the instructions here. To predict the test sentences and generate the SECS, please use the Jupyter Notebooks available here.
To fully replicate experiment 1 we provide a recipe on Coqui TTS. This recipe downloads, resample, extracts the speaker embeddings and trains the model without the need of any changes in the code.
If you want to use the latest version of the Coqui TTS you can get the config.json from the Coqui released model.
With config.json in hand, you first need to change the "datasets" configuration to your dataset. Using the config.json with the "datasets" configuration adjusted you need to extract the speaker's embeddings using our released speaker encoder using the following command:
python3 TTS/bin/compute_embeddings.py --model_path model_se.pth.tar --config_path config_se.json --config_dataset_path config.json --output_path d_vector_file.json
"model_se.pth.tar" and "config_se.json" can be found in Coqui released model while config.json is the config you set the paths for.
Other parameters that you should change are on the "config.json":
"d_vector_file": Now that you have the speaker embedding file (d_vector_file.json) adjust the "d_vector_file" parameter on the config setting to the path of the speaker embedding file.
"output_path": the path for saving the checkpoint and training logs
"speaker_encoder_config_path": The speaker encoder config to use to compute the speaker cosine similarity loss/speaker consistency loss ( set it to the config_se.json path)
"speaker_encoder_model_path": The speaker encoder checkpoint used to compute the speaker cosine similarity loss/speaker consistency loss ( set it to the "config_se.json" path)
Now that you have the config.json configured to replicate the training you can use the following command (if you like you can use the --restore_path {checkpoint_path} to do transfer learning from a checkpoint and speed up the training:
python3 TTS/bin/train_tts.py --config_path config.json
Citation
Preprint
@ARTICLE{2021arXiv211202418C,
author = {{Casanova}, Edresson and {Weber}, Julian and {Shulby}, Christopher and {Junior}, Arnaldo Candido and {G{\"o}lge}, Eren and {Antonelli Ponti}, Moacir},
title = "{YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for everyone}",
journal = {arXiv e-prints},
keywords = {Computer Science - Sound, Computer Science - Computation and Language, Electrical Engineering and Systems Science - Audio and Speech Processing},
year = 2021,
month = dec,
eid = {arXiv:2112.02418},
pages = {arXiv:2112.02418},
archivePrefix = {arXiv},
eprint = {2112.02418},
primaryClass = {cs.SD},
adsurl = {https://ui.adsabs.harvard.edu/abs/2021arXiv211202418C},
adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
Published paper at ICML
@inproceedings{casanova2022yourtts,
title={Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone},
author={Casanova, Edresson and Weber, Julian and Shulby, Christopher D and Junior, Arnaldo Candido and G{\"o}lge, Eren and Ponti, Moacir A},
booktitle={International Conference on Machine Learning},
pages={2709--2720},
year={2022},
organization={PMLR}
}
