Giters

rishikksh20 / hifigan-denoiser

HiFi-GAN: High Fidelity Denoising and Dereverberation Based on Speech Deep Features in Adversarial Networks

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Loss

saivinaypsv opened this issue · comments

commented

Hello Rishi,

I am experimenting Speech-Bandwdith-Extension(NarrwoBand - SuperWIdeBand) using this network without Post-Net. I could observe that Generator loss going high-value and that to fluctuating, But evaluating with unseen signal , I could able to recunstruct SuperWideBand from NarrowBand signal.

I am having confusion on model convergence .. Can you plz give some insights on model convergence?

commented

Steps : 162235, Gen Loss Total : 202.633, Sample Error: 0.007, Mel-Spec. Error : 0.322, s/b : 2.419 , DiscLoss : 1.287
Steps : 162240, Gen Loss Total : 376.593, Sample Error: 0.023, Mel-Spec. Error : 1.689, s/b : 2.414 , DiscLoss : 1.315
Steps : 162245, Gen Loss Total : 162.670, Sample Error: 0.007, Mel-Spec. Error : 0.345, s/b : 2.419 , DiscLoss : 1.403
Steps : 162250, Gen Loss Total : 200.409, Sample Error: 0.007, Mel-Spec. Error : 0.322, s/b : 2.421 , DiscLoss : 1.389
Steps : 162255, Gen Loss Total : 195.019, Sample Error: 0.008, Mel-Spec. Error : 0.397, s/b : 2.416 , DiscLoss : 1.355
Steps : 162260, Gen Loss Total : 355.557, Sample Error: 0.013, Mel-Spec. Error : 1.496, s/b : 2.415 , DiscLoss : 1.309
Steps : 162265, Gen Loss Total : 145.445, Sample Error: 0.007, Mel-Spec. Error : 0.365, s/b : 2.418 , DiscLoss : 1.487
Steps : 162270, Gen Loss Total : 214.779, Sample Error: 0.012, Mel-Spec. Error : 0.415, s/b : 2.423 , DiscLoss : 1.356
Steps : 162275, Gen Loss Total : 141.533, Sample Error: 0.005, Mel-Spec. Error : 0.330, s/b : 2.417 , DiscLoss : 1.451
Steps : 162280, Gen Loss Total : 121.838, Sample Error: 0.007, Mel-Spec. Error : 0.354, s/b : 2.414 , DiscLoss : 1.398
Steps : 162285, Gen Loss Total : 122.640, Sample Error: 0.007, Mel-Spec. Error : 0.347, s/b : 2.422 , DiscLoss : 1.322

commented

Hello Rishi,

I am experimenting Speech-Bandwdith-Extension(NarrwoBand - SuperWIdeBand) using this network without Post-Net. I could observe that Generator loss going high-value and that to fluctuating, But evaluating with unseen signal , I could able to recunstruct SuperWideBand from NarrowBand signal.

I am having confusion on model convergence .. Can you plz give some insights on model convergence?

Hi, Bro
Can I get a copy of your processed training data about this code, I don't understand the specific definitions of the folders and file trees in the code, such as gt_noise, gt_clean, generated, etc.