nsynth_wavenet icon indicating copy to clipboard operation
nsynth_wavenet copied to clipboard

The waveform generated by parallel wavenet looks like the result of the teacher, but quiet noisy.

Open bfs18 opened this issue 6 years ago • 9 comments

teacher waveform te_waveform student waveform st_waveform teacher spectrogram te_spec student spectrogram st_spec

The student is run with weight normalization and the default configurations in the parallel_wavenet.json. The wave generated by the student is still quiet noisy. I am testing data dependent initialization fro weight normalization.

bfs18 avatar Jun 18 '18 10:06 bfs18

Nice. How is the f0 looks like?

zhang-jian avatar Jun 18 '18 10:06 zhang-jian

hi @bfs18 ! Nice. How many steps have been trained to get that results?

maozhiqiang avatar Jun 25 '18 09:06 maozhiqiang

@maozhiqiang The above result is evaluated at 50k steps. I also generate waves at 150k steps. It is a bit clearer, but still noisy.

bfs18 avatar Jun 25 '18 12:06 bfs18

@bfs18 how about the teacher network's performance? the good teacher network is very very important!

maozhiqiang avatar Jun 26 '18 00:06 maozhiqiang

https://github.com/bfs18/nsynth_wavenet/blob/data_dep_init/tests/pred_data-pwn-failed_cases/gen_LJ001-0001-stft_abs.wav I got a bit clearer waveform. @zhang-jian @maozhiqiang

bfs18 avatar Jun 30 '18 17:06 bfs18

Just wondering if you have tried to train the student model on KL loss only?

zhang-jian avatar Jul 03 '18 10:07 zhang-jian

@zhang-jian I tried that, but didn't get meaningful result. Experimenting on KL + power loss is more promising. Besides I have limited computing resource. I didn't spent much time on it.

bfs18 avatar Jul 03 '18 12:07 bfs18

Hi @bfs18 , How many iterations does the teacher network take to get a result like tests/pred_data-use_mu_law+mol/gen_LJ001-0001.wav ?

HallidayReadyOne avatar Aug 02 '18 10:08 HallidayReadyOne

HI @HallidayReadyOne 200K steps.

bfs18 avatar Aug 02 '18 21:08 bfs18