How I could improve my model ?

[Snowad14]

I trained my french model on the M-AILABS dataset (about 200h) on 10000 step, text LR Weight on 1.

I see a clear improvement compared to my old 20k sample datasets especially since the English accent has almost disappeared but it still struggles to speak some words and the voice cloning just doesn't work (voice not really recognizable)

Do you have any ideas in my dataset or in the training that could help my model?

I have some ideas like trying another tokenizer (but I don't see much interest since English and French have about the same letter) or trying a dataset with a wider variety of speakers even if it means sacrificing some quality by using commonVoice