don't make connections

[MXC48]

the TTS is great but when he speaks in French he doesn't make the connections between words.

[csukuangfj]

the TTS is great but when he speaks in French he doesn't make the connections between words.

which model are you using?

Could you explain what "connections" means? would be great if you can post an example audio with the corresponding text.

[MXC48]

I use the UMPC medium model, but it works for all of them. In French, a liaison (connection) is the last letter of a word that can influence the next one.

A video that explains : https://www.youtube.com/watch?v=yRCD8vgohZo

[csukuangfj]

https://www.youtube.com/watch?v=yRCD8vgohZo

Thanks for the link. The video is very informative.

I think we have fixed the issue by #453

The following is a sample wav generated by the following command

python3 ./python-api-examples/offline-tts.py \
  --vits-model=./vits-piper-fr_FR-siwis-low/fr_FR-siwis-low.onnx \
  --vits-tokens=./vits-piper-fr_FR-siwis-low/tokens.txt \
  --vits-data-dir=./vits-piper-fr_FR-siwis-low/espeak-ng-data \
  --debug=1 \
  --output-filename=./siwis-fr.wav \
  "les oranges"

https://github.com/k2-fsa/sherpa-onnx/assets/5284924/cbe511be-3296-46ac-9c2d-24bb51a4488b

---

You can find the models at https://github.com/k2-fsa/sherpa-onnx/releases/tag/tts-models

[csukuangfj]

@MXC48

Please try it using our huggingface space. https://huggingface.co/spaces/k2-fsa/text-to-speech

[MXC48]

@MXC48

Please try it using our huggingface space. https://huggingface.co/spaces/k2-fsa/text-to-speech

yes that's much better thank you very much !

[csukuangfj]

Shall I mark it as fixed now?

[MXC48]

Shall I mark it as fixed now?

Yes you can !