Jing wrote:
> Hi,
> 
> I am trying to implement leaky NLMS for adaptive beamforming, but I found
> that the noise reduction performance is worse than that of NLMS. The data
> I used are simulated data, and I assume the noise and speech sources are
> well seperated. Can anybody give me some suggestions about applying leaky
> NLMS for GSC? Is Leaky really better than that of NLMS? 

Leaking will always make the adaption accuracy poorer than not leaking. 
Think about it - you keep shrinking the coefficients, so the adaption is 
always fighting to reach its target. It never quite gets there.

Leaking is a way to stop coefficients wandering off to extreme values. 
This can happen at places where the signal is always close to zero 
(think of things like the mid-point between symbols in a channel 
equalizer). No matter how far off the coefficient wanders 
big_coeff*tiny_signal is always very small, and never coaxed the 
wandering coefficient back towards zero.

If you face wandering coefficients, leaking is a huge benefit. If you 
don't, leaking is a looser. The level of leaking you should use is a 
compromise between stopping the coefficient wandering (big leak is 
better) and avoiding detuning of the adaption (small leak is better).

Regards,
Steve

Jing wrote:

> Hi,
> 
> I am trying to implement leaky NLMS for adaptive beamforming, but I found
> that the noise reduction performance is worse than that of NLMS.


Yes, this is what expected.

> The data
> I used are simulated data, and I assume the noise and speech sources are
> well seperated. Can anybody give me some suggestions about applying leaky
> NLMS for GSC? Is Leaky really better than that of NLMS? 
> 

It depends. The amount of leak is the tradeoff between the stability and 
the accuracy. You should optimize it for your application.



Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

"Jing" <jing.deng@gmail.com> wrote in message
news:9sWdnSfENbOQg1nenZ2dnUVZ_s-dnZ2d@giganews.com...
> Hi,
>
> I am trying to implement leaky NLMS for adaptive beamforming, but I found
> that the noise reduction performance is worse than that of NLMS. The data
> I used are simulated data, and I assume the noise and speech sources are
> well seperated. Can anybody give me some suggestions about applying leaky
> NLMS for GSC? Is Leaky really better than that of NLMS?
>
> Thanks a lot!
>
> Best Regards,
> Jing
>
>
>
I suspect that leaky anything won't be as good as pure integrators in an
algorithm. For stationary data NLMS woudl come out on top. Try
non-stationary data  - the leaky one may track better.


Naebad

Hi,

I am trying to implement leaky NLMS for adaptive beamforming, but I found
that the noise reduction performance is worse than that of NLMS. The data
I used are simulated data, and I assume the noise and speech sources are
well seperated. Can anybody give me some suggestions about applying leaky
NLMS for GSC? Is Leaky really better than that of NLMS? 

Thanks a lot!

Best Regards,
Jing