Open ScottishFold007 opened 1 month ago
Interesting, can you share samples ? Probably vocos is not adapted. I found it to generalize poorly on languages that are not english. Switching to a more robust representation, and also setting an higher bitrate would help significantly.
I trained some Chinese models with this work of yours and made a discovery: The emotional pronunciation is quite good, but there is a lack of high-frequency information. The issue is that too much information is lost due to token loss, so there is a need for flow matching(eg,Matcha-TTS) to compensate for some of it. What do you think?