tokenizer

#3
by yeserumo11 - opened

hey bro, what the tokenizer you use for rus, Is it the default tokeniser provided by mrq?

Hi, @yeserumo11 ! Yes, I didn't do any changes to the tokeniser. My understanding is that fine-tuning wouldn't really be possible if I changed the tokeniser. The original can already produce results in Russian, though with a strong accent.

thank you, one more question, im not sure if im understandling this correctly, the training audio is single-speaker audio clip, so when generate with the trained model, it can generate the voice like the guy and speak textual content that has not appeared in the training data. But it may performace bad if i want to generate another person's voice using this model, is that right? So i should train a model for every single speaker.

thank you, one more question, im not sure if im understandling this correctly, the training audio is single-speaker audio clip, so when generate with the trained model, it can generate the voice like the guy and speak textual content that has not appeared in the training data. But it may performace bad if i want to generate another person's voice using this model, is that right? So i should train a model for every single speaker.

yeserumo11 changed discussion status to closed
yeserumo11 changed discussion status to open

thank you, one more question, im not sure if im understandling this correctly, the training audio is single-speaker audio clip, so when generate with the trained model, it can generate the voice like the guy and speak textual content that has not appeared in the training data. But it may performace bad if i want to generate another person's voice using this model, is that right? So i should train a model for every single speaker.

no , because
"Tortoise is a text-to-speech program built with the following priorities:

you can see it support strong multi-voice speaker
In case your model run wrong, can't clone another voice, maybe you training model wrong way

Sign up or log in to comment