fix "." (dot) pronounced #308
Conversation
|
Thank you for the PR! I'm not sure yet if this is a good idea, I'm afraid it might cause unexpected issues. Perhaps a safer approach would be to only remove periods at the end of the input? |
|
according to our hundreds test from more than 100 languages, and thousands users reports, I can say removing the "dot" is worse , causing change of the sentence voice tone confusing its sense and hallucinations sometimes. |
Should this be not removing it is worse? On the other hand in #317 (comment) you suggested adding punctuation can help with short utterances, so removing it always might not be the best?
There is probably no clean way to fix it because it's likely due to some issue in the original XTTS training, which is not reproducible. |
no, removing it is worse, not worth! :) now if you really want more and more report of user complaining about the "dot " be pronounced (in all languages btw). so it's up to you. on our side, we prefer for now use our patch to avoid that. |
| @@ -258,10 +258,11 @@ def expand_abbreviations_multilingual(text, lang="en"): | |||
| ("$", " dollar "), | |||
| ("£", " pound "), | |||
| ("°", " degree "), | |||
| (".", ""), | |||
There was a problem hiding this comment.
no, removing it is worse, not worth! :) the dot should stay with a space between the word and the dot. then use my patch and everyone is happy to have a TTS not pronouncing "dot" at each sentence.
Ok. I'm sorry, but I don't fully understand your change yet then. With these lines, you are removing the dot, no?
|
@ROBERT-MCDOWELL , with Italian model, it is already pronouncing "." as punto. I haven't seen this behavior with other models. |
|
it's only with xttsv2 |
TODO: it needs to check if the "." dot is between numbers so it has to be pronounced (num2words?)