search

rishikksh20 / hifigan-denoiser

HiFi-GAN: High Fidelity Denoising and Dereverberation Based on Speech Deep Features in Adversarial Networks
Apache License 2.0
208 stars 45 forks source link

Data simulation and augmentation #5

Open v-nhandt21 opened 2 years ago

v-nhandt21 commented 2 years ago

Can you detail the way you are using to make the noise audio for training?

Does it the same with described in the paper?

image

Are you using kaldi or any tool for this, and can you share your noise dataset !

Thank rishikksh !

LXnn058 commented 2 years ago

Hi, have you solved this problem?

v-nhandt21 commented 2 years ago

Hi, have you solved this problem?

With my experiment, the data augmented with Kaldi give better results than some methods that augment on the fly

LXnn058 commented 2 years ago

Are the parameters the same as those set in the paper? Can you provide me the corresponding processing script? thanks.

v-nhandt21 commented 2 years ago

I think the parameter depending on the test set / practical problem you want to solve.

My config is SNR between 10dB and 30dB, use MUSAN as additive noise

You can check out the public source code here: https://github.com/zhaoyi2/audio_augment

LXnn058 commented 2 years ago

This helped me a lot, thanks！

skol101 commented 2 years ago

@v-nhandt21 I simply added audiomentations (RoomSimulator, Backgroundnoise) and stored to separate directory, which I use in "input_wavs" params after step for postnet training (250k in my case, though it's 500k in the default config).

v-nhandt21 commented 2 years ago

@v-nhandt21 I simply added audiomentations (RoomSimulator, Backgroundnoise) and stored to separate directory, which I use in "input_wavs" params after step for postnet training (250k in my case, though it's 500k in the default config).

Yep, it seems hard for the model to converge on the fly augmentation

velonica0 commented 2 years ago

Hi, have you solved this problem?

With my experiment, the data augmented with Kaldi give better results than some methods that augment on the fly

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.