The synthesized audio sounds like it through the voice changer. about multilingual_text_to_speech

[Tomiinek]

Hello, glad to hear that your two-language model is working 👍

Again, it is hard to tell what is happening there. I haven't encountered this problem and have no idea what could be the cause.
I would guess that your training recordings are not normalized properly.

[Tomiinek]

Could it be because of the "speaker_embedding_dimension": 32 ?

I do not think so. The model just cannot speed up output recordings. I do not know what is it caused by, but it seems like you are feeding there faster training examples (of a part of your dataset).

[michael-conrad]

A shot in the dark. You might try changing the phonemizer call to include stress marks in it's output and create new training metadata from scratch with the revised phonemes and see if that helps.

[YihWenWang]

Do you mean "ffmpeg normalization" ?
Could it be because of the "speaker_embedding_dimension": 32 ?

[YihWenWang]

Thank you very much.
I have solved this problem when I use the same sample rate 22050 kHz in each language dataset.

[nltoan]

can you share your config for 2 dataset? @YihWenWang
my mail is: [email protected]