IPP update 4 echo canceller/denoiser issue

[evgeny777]

Hi, Everyone!

I'm using IPP echo canceller/denoiser in our VoIP product and it works quite good with IPP 6.1.2

After linikng with IPP update 4 I noticed that output voice signal volume became much lower if compare

with update 2. When I switched back to 6.1.2 things went back normal. What can be the reason?

[Ivan_Ryzhachkin__Int]

Hello Evgeny,
Noise reduction mode was changed from Medium to High.Much noise eliminated, and voicesignal power lowered accordingly as well.
Regards,
Ivan

[evgeny777]

Hi, Ivan!

Thanks for an answer.
Is there any way to set it back to medium level? Or I should lower noise reduction (nr) level (now nr=4).

And another question: we've recently checked echo cancellation on Dell notebook with Intel Core i5 processor and IDT HD Audio sound card. The AEC doesn't work at all if adaptation is set to AD_OFF (can hear every word spoken). If I set adaptation to AD_FULLADAPT the AEC works better but still can hear echo (however it is softer). Strange but on most of test machines AD_OFF works very well. What actually AD_FULLADAPT does? Are there any hints to improve EC?

[Ivan_Ryzhachkin__Int]

Hi Evgeny,

if nr=4 used inboth casesthenI'd like to have an answer for the followingquestion:

- is the output noise level the same for 6.1.2 and update 4?

Suppose the noise level is lower with update 4. Why not to leave as is(gaining output if needed)?

Could you please clarify upon which library was engaged for theCore i5?

regards,
Ivan

[evgeny777]

I didn't feel difference in noise level after brief test. However background noise was npt very high so I can't judge which one is better. I use Speex AGC (same thing as ALC in IPP) which doesn't work well with 6.1.4. May be switching to IPP ALC is a good alternative.

I used IPP 6.1.2 when testing on Core I5. Does 6.1.4 offer any improvement?

[Ivan_Ryzhachkin__Int]

Hi Evgeny,

There are one more differencebetween 6.1.2 and update 4. NR was changedfor not to catch noise

changing rapidly(by magnitude, which is close to voice behaviour) to avoid misdetects.

Core i5 IPP tested well. NeverthelessPX (C-not optimized) IPP versioncould be engaged to feel the

difference (leave px libs (dll's) in the project (folder) only).

IPP ALC could be tried.

Adaptation could be switched off when echo path is stable and coefficients were trained well before. But I

see theoreticalsituation when AEC works well with NR, NLP switched onand adaptation

switched OFF initially.

Regards,
Ivan

[evgeny777]

Hi, Ivan!

I finally made IPP EC work on Core i5 after switching off all audio enchancements in IDT audio driver settings. I think (although not sure) that this is because of some non-linear processings done by IDT driver
which prevent IPP EC from working normally (currently I pass sampes to EC which are untouched by audio hardware)

I will try using Vista/Windows7 WASAPI loopback recording feature to pass samples captured directly from audio card to echo canceller. Hope this will help. Thanks for detailed explanation

Eugene.

[Ivan_Ryzhachkin__Int]

Hi Evgeny,
Sure, any distortions arebad for any AEC.
I'm glad for you.
Thank you for positive feedback!
Ivan