Fine-Tuned Toucan for Hindi : Fails to Identify Sentence Boundaries

[anjalyv]

Hello,

I am attempting to use Toucan for Indian languages, finetuned model for Hindi. The sentence endings are not captured correctly in the generated audio. In Hindi, we have a different sentence-ending symbol (|) instead of a period (.). Could this be the reason? Is there any way to improve this?

Thank you for your incredible work.

[Flux9665]

Hi! Sorry for taking so long to reply, I was overwhelmed with things to do recently. Yes, your suspicion is correct. I was not aware that Hindi uses a different sentence-ending symbol. Now it should be better during inference. You might want to run the feature-extraction stage and the finetuning again. The fix will affect all phoneme sequences that are generated from now on, but in the cache files from your previous run, there are still the incorrect input sequences, where the | is removed, because the same unicode character is also an unsupported modifier of the phonetic alphabet.

[jogong2718]

Hi I think i saw the fix to this issue in this commit: (fa64ca6). There is a trailing comma at the end of the line that is causing some problems for me as it turns the string into a tuple is this intentional?

[Flux9665]

good catch, I don't know how I didn't notice that. It is fixed now, thank you!