DigitalPhonetics / IMS-Toucan

Controllable and fast Text-to-Speech for over 7000 languages!
Apache License 2.0
1.47k stars 166 forks source link

Inference speed #185

Closed a779159990 closed 3 months ago

a779159990 commented 3 months ago

Hello, your job is great, but I found that some languages have a much faster reasoning speed than most languages during the reasoning process, such as Japanese. Is this problem normal?

Flux9665 commented 3 months ago

Thank you! The reason for the speed difference between languages is in the preprocessing. The model does the same for all languages, but it requires transcriptions in the phonetic alphabet. For some languages it takes longer to convert the regular written form to the IPA form. For Japanese, we use PyKakasi to unify the different writing systems and then use Transphone to get IPA symbols. Transphone is a bit slow.

a779159990 commented 3 months ago

I understand, thank you!