SpeechT5-base-cs-tts

This is a monolingual Czech SpeechT5 base model pre-trained from 120 thousand hours of Czech speech and a large Czech text corpus with 17.5B words. It has been introduced in a paper Zero-Shot vs. Few-Shot Multi-Speaker TTS Using Pre-trained Czech SpeechT5 Model accepted to TSD2024 conference. It is meant to be used as a starting checkpoint for Czech TTS fine-tuning.

Important note: This is only a pre-trained model, so it has not yet been trained for the TTS task.

Examples of fine-tuned voices are available at https://jalehecka.github.io/TSD2024.

See our paper for details.

Paper

The pre-print of our paper is available at http://arxiv.org/abs/2407.17167.

Citation

If you find this model useful, please cite our paper:

@inproceedings{lehecka2024multispeakertts,
  title = {{Zero-Shot vs. Few-Shot Multi-Speaker TTS Using Pre-trained Czech SpeechT5 Model}},
  author = {
    Jan Lehe\v{c}ka and
    Zden\v{e}k Hanzl\'i\v{c}ek and
    Jind\v{r}ich Matou\v{s}ek and
    Daniel Tihelka
  },
  booktitle = {{Text, Speech, and Dialogue}},
  publisher = {{Springer International Publishing}},
  year = {2024},
  note = {In press},
  url = {https://arxiv.org/abs/2407.17167},
}

Usage

This is a TTS variant of the SpeechT5 model, i.e., the input modality is text (processed via the text pre-net), and the output modality is speech (processed via the speech post-net). The model has the same format as the microsoft/speecht5_tts compatible with the SpeechT5ForTextToSpeech class.

In order to use this model for text-to-speech, it must be fine-tuned on labeled TTS data.

The usage is the same as for the microsoft/speecht5_tts

Related works