search

NVIDIA / tacotron2

Tacotron 2 - PyTorch implementation with faster-than-realtime inference
BSD 3-Clause "New" or "Revised" License
5.11k stars 1.39k forks source link

How to set Tacotron2 for Polish language? #468

Open VexorionReal opened 3 years ago

VexorionReal commented 3 years ago

Hey! I am new here! Do you know how to set Tacotron 2 for Polish language samples?

41WhiteElephants commented 3 years ago

Yes, I have succesfully trained Polish language on Czubówna speech samples. 3-6 hours of dataset would be fine. The best way is to resume training with your polish dataset on pre-trained english model. After about 30k of iterations you should see comprehendable results. Add polish symbols to https://github.com/NVIDIA/tacotron2/blob/master/text/symbols.py & then you should start new training with polish dataset with flags --checkpoint_path official_english_model.pt and --warm_start

here samples from my polish model: https://drive.google.com/file/d/1SvsbdRcw5bADOSyh6YM2A-6u_80veH6d/view?usp=sharing

rave974 commented 3 years ago

followed, train seems to wotk but on inference i get error that says alphabets dont match

41WhiteElephants commented 3 years ago

Hey, take a look at my fork of tacotron, it's the one I used for training with all symbols and weights adjusted for polish speech ;) https://github.com/41WhiteElephants/tacotron2

rave974 commented 3 years ago

works, i don't know why i had error on inference, but i reinstalled everything and followed your steps and it works. non english dataset with 3700wavs trained 100epochs with continue train from english checkpoint. inferencing using pretrained english waveglow gives perfect voice sound quality, but intonation, speed, punctuation, and some words sometimes are not good enough i guess it needs more training or a bigger dataset.

NeonBohdan commented 2 years ago

Very cool @41WhiteElephants can you please tell where the Czubówna speech samples can be downloaded