yatingmusic / ddsp-singing-vocoders Goto Github PK

How to change mel parameters including num_mel_bins, hop_size, win_size, fmin, fmax to fit my own acoustic model?

HI,
I found that in the vocoder.py:134, there are two parapmeters "hz_min" and "hz_max", are they the same with the preprocess.py: line 62, 63 mel_fmin, mel_fmax? why hard coded?

Thanks for the good job.
We are finding a vocoder for TTS, we tried the SawSingSub, our audio sample rate is 16000, to cooperate with our acounstic model, I had changed the params in preprocess.py, set hop_length=200, win_length=800, n_mel_channels=80, I also changed the , sawsingsub.yaml, accordingly , set block_size = 200, left other settings unchanged.

I have trained the model for nearly 1 million steps, but the quality of generated waveform is far worse then that of HiFi-GAN with the same steps, Is there any parameter adjustments I had missed?

The Blured Spectrogram of SawSingSub generated speech

thanks for the good job.
I am trying to run an example training process, but it can not run on my pytorch version 1.7.1. some functions are not there.

eg. ddsp/core.py:264 torch.fft.irfft()

how to adjust the project to run in my old torch?  it seems that the torch.irfft is not the answer.