Tacotron2: how to improve the quality of male's voice?

[wotulong]

Thanks for your job.I have trained your code on my own dataset, as a result .I can get a good quality of voice when I train with female's voice, but get a poor quality of voice when using male's voice to train( both source voice have a very good quality, male 's data is easy to alignment when set tacotron_output_perstep=1, but male's data is hard to alignment). by the way ,I keep the hparams same as yours. Do you have any suggestions for me? Thank you.

[akshitmittal1]

what are issues are there with the male audio quality? When listening output audio, can we understand the words or we can't understand the words at all?

[wotulong]

what are issues are there with the male audio quality? When listening output audio, can we understand the words or we can't understand the words at all?

Yes, we can understand every word in the voice.The male's voice have more reverb than female's.

[wotulong]

what are issues are there with the male audio quality? When listening output audio, can we understand the words or we can't understand the words at all?

Yes, we can understand every word in the voice.The male's voice have more reverb than female's.

and not as cleary as female's. sound like have some noise.

[akshitmittal1]

Okay, try to infer it at a larger sampling rate. That thing worked for me, it increases the pitch of the voice and makes it clearer.

[wotulong]

Okay, try to infer it at a larger sampling rate. That thing worked for me, it increases the pitch of the voice and makes it clearer.

Thanks.My samplerate is 22050, what's yours?

[akshitmittal1]

It all depends on your training data as well as the hyperparameters. Did increasing sampling rate while infering work for you?

[rafaelvalle]

Please note that increasing the sampling rate should not alter the pitch of the voice nor necessarily make it cleaner. Increasing the sampling rate makes it possible to represent higher frequencies. One
can represent frequencies up to half the sampling rate. For eample, in 16khz one can represent frequencies up to 8khz.

[rafaelvalle]

Closing due to inactivity.

[Approximetal]

I also have this problem as above said "The result get a poor quality of voice with reverb and sound like have some noise." It has trained 450k, and loss decreases slower now, should I continue training ? What can I adjust? Any other mothods could I try to improve the qulity? @rafaelvalle

[Liujingxiu23]

I have the same problem as you @wotulong , male voice worse than female, noise if male waves synthesized is perceptible, while these is no noise of female wavs. Have you find any solution or any reason?

[wotulong]

I have the same problem as you @wotulong , male voice worse than female, noise if male waves synthesized is perceptible, while these is no noise of female wavs. Have you find any solution or any reason?

I think maybe it because the voice of male has more energy in low freq, you can try decrease the value of "fmax" in hparams file.

[ksaidin]

Check this issue and it's solution. It is suggested that training waveglow with your own dataset produces original like output.

[tugstugi]

@wotulong which "fmax" have you tried for your male voice?

[wotulong]

@wotulong which "fmax" have you tried for your male voice?

fmin=60, fmax=7000.