search

egorsmkv / ukrainian-tts-datasets

🇺🇦 Open Source Ukrainian Text-to-Speech datasets
Apache License 2.0
12 stars 1 forks source link

Test with dataset tetyana #2

Closed magicse closed 3 months ago

magicse commented 3 months ago

Training test with dataset tetyana without correct stress in words. ukr_test_40_28.zip

egorsmkv commented 3 months ago

It's intelligible, what do you use to make it?

magicse commented 3 months ago

StyleTTS and the size of the model came out to be about 311 MB

patriotyk commented 3 months ago

@magicse Do you speak ukrainian? For what reason do you train it?

magicse commented 3 months ago

Yes. I try to get more quality of voice. And may be in future try port it to embedded devices like orange Pi for local voice assistant.

egorsmkv commented 3 months ago

@magicse then join our community here - https://t.me/speech_synthesis_uk

magicse commented 2 months ago

Hi @egorsmkv Test with StyleTTS https://huggingface.co/spaces/tracert/ukr-tts size of model 160 mb only. I use espeak- ng for phonemisation and stress... but every new version of espeak has problems with Ukrainian language, because someone make incorrect changes of rules and dictionary in espeak github repo. And my local espeak and hugginface espeak work with different rules ))) That is why I want try to use ipa-uk and ukrainian-word-stress but there are some problems with it. https://github.com/lang-uk/ukrainian-word-stress/issues/7#issue-2551661934 The main task is to get a high-quality voice with a small model size. After that, I will try to transfer it all to cpp. I have already moved Hifi-GAN to cpp https://github.com/magicse/ncnn-hifi-GAN