XTTS pronounces punctuation

[airg14]

Hi everyone,

I'm a french user and i'm trying tts_models/multilingual/multi-dataset/xtts_v2 model which is very useful. To explore all voices I made a recorded the same text for everyone. And something odd happens : speakers pronounce punctuation, especially periods. Is there a way to avoid this and made the model treat punctuation like a pause ?
My exemple in french :

Pour créer une ambiance intemporelle et immortaliser ce moment, nous vous invitons à puiser votre inspiration dans les années çà ! Que vous soyez plutôt Tupac Shakur, les Spice Girls ou Nirvana, à vous de choisir. Si vous préférez l'univers cinématographique, vous pourrez opter pour des inspirations telles que Thelma & Louise, Les Affranchis ou Matrix. Laissez libre cours à votre créativité et surprenez-nous.

Thanks for your help !

[JeanBus]

Hello airg14,

I'm having the same issue : instead of interpreting the punctuation, it "reads" it.
Is there any parameter to set to change that ? Can we at least tell the speaker to mark au pause instead of saying "comma", for exemple ?

Thanks,

[eginhard]

Yes, this is a common issue with XTTS in all languages and can't be fully avoided. The best workaround is to synthesize one sentence at a time and remove punctuation from the input.