porting the dithering function from kaldi

[KarelVesely84]

No description provided.

[KarelVesely84]

okay, I added the missing part, the local build including the test binaries was okay...

[KarelVesely84]

the failed linux-macos test seems unrelated

there is a return at unusual place in that failing test test_rfft.py

[KarelVesely84]

for me it is ready, i just changed the default dithering constant

k2 uses audio samples [-1..+1]
kaldi was using audio samples [-32k..+32k], hence the constant is 2^15x smaller

[csukuangfj]

Thanks! Will check it tomorrow morning.

[csukuangfj]

Thank you for your contribution!

[nshmyrev]

Hey, can we please change dither constant to be 1.0 like it was before. 0.00003 vs 1.0 can cause a lot of confusion in existing code.

[csukuangfj]

Could you create a PR to change it to 1.0?

[KarelVesely84]

The problem is, that in some situations 1.0 is the correct value (for +/- 32k audio signal), while in other situation 0.00003 is the correct value (+/- 1.0 audio signal range), and both options are possible use-cases. So there will be confusion either on one or the other side. And the detection cannot be 100% reliable, as there can be an input signal with hard zero's in it.

So it is better to document it, and give a hint to the users.

But, it is true that if default is 1.0, being internally divided into 0.00003. And, occasionally it is used with +/-32k input singal, it will effectively disable the dithering. While as it is now, using 1.0 will break the system with +/-1.0 audio signal. So, changing the default to 1.0 is a safer way...

[nshmyrev]

Let it stay then, sorry for confusion and thanks for the patch.