A little gap between this repo and the official result.

[TeaPoly]

Thanks for this great repo.
I have dump audio from https://www.amazon.science/blog/how-amazon-chimes-challenge-winning-noise-cancellation-works.
percepnet-noisy.wav and percepnet-clean.wav is probe and target audio. percepnet.oracle_bands.wav is process only using band from Amazon. The result is very close to your repo test-bands.pcm. But the band gain and pitch strength result percepnet-oracle-pitch.wav from Amazon is different to the result test-pitch.pcm from your repo. It sounds more lower volume and has a little roughness.
Maybe we can find clues together.

Here is all audio files:
https://drive.google.com/file/d/1Kf6R0nOGI1LyLZ1wDQlz90LXSXyr92AN/view?usp=sharing

[jzi040941]

Hello, Thank you for reporting the difference between result from my repo and offical Amazon result.
I've test few things and finally get the most similar output at branch filterstrength
currently, it's on pull requests
please let me know your opinion for this version

[jzi040941]

https://drive.google.com/file/d/1VcqJDZt2XkAhbEdl348baXfI6GYzdjB4/view?usp=sharing

this is the output of branch filterstrength from the input files that you share.

[TeaPoly]

Great job. It sounds good.

[TeaPoly]

Have you test DNS 2020 synthetic test set (no_reverb/with_reverb) PESQ (WB/NB) with your repo?

[TeaPoly]

Here is no_reverb synthetic test set PESQ1(ITU-T P.862.1), PESQ2(ITU-T P.862.2) result from this repo frontend with an oracle (ideal ratio model) result.

Baseline result from Real-time Denoising and Dereverberation with Tiny Recurrent U-Net.

			no_reverb
	Size(M/MB)	RT	PESQ1	PESQ2
Noisy	-	Yes	2.45	1.58
NSnet	1.2/4.84	Yes	2.68	1.81
ConvTasNet	5.08/19.38	No	-	2.73
PoCoNet1	50/190.73	No	-	2.71
PoCoNet2	50/190.73	No	-	2.75
DCCRN-E	3.7/14.11	Yes	3.27	-
DCCRN-CL	3.7/14.11	No	3.26	-
TRU-NET(FP32)	0.38/1.45	Yes	3.36	2.86
TRU-NET(INT8)	0.38/0.36	Yes	3.35	2.84
Percepnet-Oracle	-	-	3.37	2.60

[TeaPoly]

I think post-filtering maybe degrade PESQ2 result. I will try more experiment.

[TeaPoly]

Add Percepnet-Oracle without post-filtering result.
Post-filtering will get degraded PESQ result which mentioned in original paper.

			no_reverb
	Size(M/MB)	RT	PESQ1	PESQ2
Noisy	-	Yes	2.45	1.58
NSnet	1.2/4.84	Yes	2.68	1.81
ConvTasNet	5.08/19.38	No	-	2.73
PoCoNet1	50/190.73	No	-	2.71
PoCoNet2	50/190.73	No	-	2.75
DCCRN-E	3.7/14.11	Yes	3.27	-
DCCRN-CL	3.7/14.11	No	3.26	-
TRU-NET(FP32)	0.38/1.45	Yes	3.36	2.86
TRU-NET(INT8)	0.38/0.36	Yes	3.35	2.84
Percepnet-Oracle	-	-	3.37	2.60
Percepnet-Oracle(no-pf)	-	-	3.58	3.03