autumn owl
I'm working on training a voice model using only non-English audio clips from native speakers of other languages. My goal is to use voice conversion to have them speak English fluently, but I’m concerned the output will have a foreign accent since no English phonemes were present in the training data.
Is there a way to get a native-sounding English output from a model trained only on non-English data?
Thanks in advance.
sterile ice
having non-english phonemes in the training data is not much of a problem, having no english phonemes is a problem
setting index to 0, or borrowing an index from an english model, may fix the accent problem
sterile ice
That doesn't seem like a good idea
but it works
russian voice model with english audio
autumn owl
Hmm
But how does that work? Would I have to find some English model of a voice that sounds similar to the voice I'm training?
sterile ice
model.pth is the voice, model.index is phonemes
during inference the process extract phonemes from the audio and tries to find something that matches in the index
the accent comes from finding something that is way way off the target because your training data does not have the required phonemes
verbal olive
I'm working on training a voice model using only non-English audio clips from na...
it depends much, afaik japanese models mostly struggle on that despite trying to use index file of a native english speaker