arabic-tacotron-tts icon indicating copy to clipboard operation
arabic-tacotron-tts copied to clipboard

Published 20 hours ago • verified publisher icon youssefsharief

No alignment for Urdu

Open AownMohammad opened this issue 5 years ago • 6 comments

Hi, I have used this model to train on my speech Urdu dataset. It contains 10000 .wav files of 15 hours speech dataset. Average file is size 5.4 sec. I used default parameters and trained for 50000 steps. I used transliterated labels and transliteration_cleaner. Here is the alignment. step-50000-align

AownMohammad avatar Feb 23 '20 19:02 AownMohammad

Can you please share the synthesize audio file .

tayyabvohra avatar Feb 26 '20 09:02 tayyabvohra

Thanks for replying. Here are the synthesized audio files at 50000 steps. www.aown.me/eval-50000-0.wav www.aown.me/eval-50000-1.wav www.aown.me/eval-50000-2.wav Here is training sample www.aown.me/45.wav

AownMohammad avatar Feb 26 '20 14:02 AownMohammad

@AownMohammad the same problem has occured with me I have iterate over 500K steps and my loss is 0.06 but still the problem is same.

tayyabvohra avatar Feb 27 '20 08:02 tayyabvohra

@tayyabvohra I have been trying with different settings but none is working. If it the model can work on Arabic then it should work on Urdu too.

AownMohammad avatar Feb 27 '20 08:02 AownMohammad

@AownMohammad I think we should tune the hyper parameters according to it.

tayyabvohra avatar Feb 27 '20 08:02 tayyabvohra

@tayyabvohra But how? training takes alot of time. Training again and again with different hyprams would become impossible.

AownMohammad avatar Feb 27 '20 10:02 AownMohammad