search

ruizhecao96 / CMGAN

Conformer-based Metric GAN for speech enhancement
MIT License
309 stars 60 forks source link

SSNR is 10.4 #23

Open DingChun-code opened 1 year ago

DingChun-code commented 1 year ago

I used your open-source code in github, ran evaluation.py, and also downloaded VCTK+DEMAND data. The model in it uses the model checkpoint you left on github in best_ckpt folder. Because the source data is 48000hz, I added the resampling part in the code. After running evaluation.py, the SSNR of the result is only 10.4, and other indicators are also slightly lower than the paper. If you have any opinions, I would appreciate it very much.

utyanndesu commented 1 year ago

We encountered the same problem, with the SSNR indicator only 10.4 and each indicator slightly lower

BancoLin commented 1 year ago

report my results:

In the following metrics the first numbers are from the TABLE I of the paper (https://arxiv.org/pdf/2209.11112.pdf), and the second ones are my results. All the results are similar except the SSNR.

PESQ: 3.41 --> 3.41 CSIG: 4.63 --> 4.67 CBAK: 3.94 --> 3.91 COVL: 4.12 --> 4.14 SSNR: 11.10 --> 10.68 STOI: 0.96 --> 0.96

despite the inconsistent numbers between code and paper, i'm quite satisfied the performance with our own data.

SherifAbdulatif commented 1 year ago

We got some complains about SSNR and it is currently being checked.

moshengmao commented 11 months ago
  • use ffmpeg to downsample the source data from 48khz to 16khz

I want to ask if I use the VCTK-DEMAND dataset to train, whether I need downsample the trainset to 16khz first or input the original data.

SherifAbdulatif commented 11 months ago

You need to downsample the data first.

On Fri, 17 Nov 2023, 2:39 am moshengmao, @.***> wrote:

  • use ffmpeg to downsample the source data from 48khz to 16khz

I want to ask if I use the VCTK-DEMAND dataset to train, whether I need downsample the trainset to 16khz first or input the original data.

— Reply to this email directly, view it on GitHub https://github.com/ruizhecao96/CMGAN/issues/23#issuecomment-1815610906, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKAXVH5UI7QPAJZVNRVY54TYE253TAVCNFSM6AAAAAAVNOHMV2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMJVGYYTAOJQGY . You are receiving this because you commented.Message ID: @.***>