Dear All, I am new to tts and deep learning. I have tried to train the tacotron2 with a small (1.5 hours) Sinhala (Sri Lanka) dataset and I think I am seeing overfitting as expected. The results are attached below. I am training with the phenomes. The final voice synthesis seems to be going in the right direction - some letters are understandable - but gibberish
I am working on creating a bigger dataset. Meanwhile is there something I can try to get better results with the data I have?
things look good to me. How is the final result ?
The final voice synthesis seems to be going in the right direction - some letters are understandable - but gibberish
I also tried the glow-tts training - since the training is 10,000 epochs as opposed to just 1000 in tacotron it seems to be much slower - is there a way to prevent running the evaluation at the end of each epoch?
run_eval in config.json
I make similar experiment with Taiwanese.
While the vocoder part with only two hours sounds great by self, the tacotron text to mel part is bad. think not enough data.
if you share more info like your config, tensorboard etc. we can maybe help
@erogol is there a way to reduce the number of parameters/layer sizes in tacotron so it will not overfit small datasets?
I am making the bigger dataset now for our language, but it takes time.
yes you need to set layer sizes manually in the source.
