Hifigan hyper params and training
Closed this issue · 4 comments
skol101 commented
I've started training, but validation melspec error doesn't really decline.
I've noticed params for 24k are quiet different from the ones used here https://github.com/kan-bayashi/ParallelWaveGAN/blob/master/egs/vctk/voc1/conf/hifigan.v1.yaml
"num_mels": 80,
"num_freq": 1025,
"n_fft": 1024,
"hop_size": 256,
"win_size": 1024,
"sampling_rate": 24000,
skol101 commented
But I can already hear that 20k sounds much better than 10k, so I guess subjective quality is more important than loss_g_melspec
Update after 30k the eval loss is going down and at 200k the quality is good.
skol101 commented
But still would like to hear why hyper params are diferrent for 24k sampling
fss1t commented
Hyperparameters conform to https://github.com/jik876/hifi-gan.
skol101 commented
Thanks