Reproduce on Voicebank dataset

[trfnhle]

First of all, thank you for your great work
I tried to reproduce on the Voicebank dataset with your code but got some problems. I try inference on checkpoint 100k but the result is not compared to your sample files and still remains background noise.

Some steps I do:

• Preprocessing Voicebank dataset with flag se
• Training without any modification
  And here is my loss figure:
  

Could you get some insight into what possibly was I doing wrong?

[neillu23]

Dear @l4zyf9x,

We found there are some issues with different PyTorch and torchaudio versions and we will also try to fix this issue soon.
Could you try with pytorch1.8.0/torchaudio0.8.0 or pytorch1.8.1/torchaudio0.8.1?

Here is my training loss figure:

Thank you!

[trfnhle]

@neillu23 Thanks for your quick response
I will try with torch and touch audio version you suggest
Btw, I have some more questions. I noticed that in sample audio, there are *raw_enhanced.wav and *enhanced.wav. What difference between them?
One more thing, when we use flag se_pre, it seems to use clean audio to condition on diffusion step. I just don't see the motive why do you use clean audio in the diffusion step

[neillu23]

Hi @l4zyf9x , sorry I missed your last message.
I've replaced torchaudio.load_wav() with the torchaudio.load() function in the new commit. You can try it with the new torch and touchaudio versions.
The *enhanced.wav are further combined with a noise signal with a ratio of 0.2 to recover high-frequency speech
as described at the end of Sec. 4.1, while *raw_enhanced.wav is the result of no combination.
The "se_pre" step was designed for our previous work DiffuSE, we tried the same initialization for CDiffuSE while writing the paper. Afterwards, we found that the pre-training step was no longer needed in CDiffuSE, since the CDiffuSE initialized randomly performed as well as CDiffuSE initialized from pre-trained parameters.
Please try the new code and let me know if you have any further questions!

[KarsonYu]

Hello,have you try the version author mentioned?And how is the performance?

[Charizard-007]

Hello,have you try the version author mentioned?And how is the performance?

Hello,have you try the version author mentioned?And how is the performance?