Issue after '.' (period), tts. voice is changing

[Chillae]

Hello everyone, great project!
I have run into an issue, sometimes after a . (period) or comma the tts switches voice. For example, I have set my voice to be jlaw, but after a . (period) it switches to a voice I can only describe as joe rogan. Which is strange, because he doesnt seem to be in the voices folder.

A work around is to not use periods, but this makes the voice sound a bit strange, without any stops.

Is there another way to add pauses to the output speech?

[neonbjb]

Someone else pointed this out to me in a different context a couple of days ago. Rogan's podcast was part of the training dataset and it would appear that I did a poor job splitting it up such that there must be some clips where both him and a target voice were talking in the same clip. This must have translated to the model learning that sometimes a period means "inject Joe Rogan talking".

Unfortunately there's not much I can do now, other than re-train the model. If I do give that a go someday, I'll certain do another filtering pass, or just remove the Rogan podcast altogether..

[Chillae]

Haha, Yes I thought the model might have been trained on joe rogan, hence his voice being there. Made me laugh when I heard it.

Is there any way to create a pause mid sentence? I guess I could re-run the sections that have changed and then edit them together, if I wanted a long sentence.

[neonbjb]

Is it really that common? I don't recall seeing this at all when I was testing the jlaw voice when I originally released Tortoise. I would have assumed this happens <10% of the time and the answer would be to "just re-render that clip".

I'm guessing you already tried using other punctuation like semicolons or dashes?

[neonbjb]

I wonder if this is caused by CLVP2, hence the recent reports..

If you have time and can easily reproduce this, would you mind reverting to fda5130 and re-trying your tests?

[scott-vector]

This doesn't work when moving back to CLVP. When we run sentences with . ; - we get a different sounding voice for each sentence, most of the time. When we remove these separators, we get a steady, single voice.