Line echo cancellation - conditions for freezing adaptation

Hi com.dsp-ers,

I know that there are some scenarios in which a line echo canceller
should suspend the process of adapting its coefficients:

1. Double talk (the most obvious case)
2. Narrow band reference signal detected
3. Numerical overload detected

I wanted to broadcast a call for ideas for detecting scenarios 2 and
3. For 2, I am currently thinking of putting in an 8 channel filter
bank and using the relation between a short term estimate of power
recorded at the output of the 8 channels to infer narrow band signals
and freeze adaptation. For numerical overload, short term estimate of
reference signal power is calculated and adaptation frozen in case a
defined upper limit is crossed. So, what I would like to ask is:

1. Are there any obvious deficiencies (that I am overlooking) in my
approach.
2. Is there a better/standard way of doing either of the two tasks? (I
could not find much in the several EC papers I went through, but these
were mainly pretty old papers by pioneers of EC).

Any help is greatly appreaciated.

Regards,
Ljubisa Bajic

eternal_nan@yahoo.com wrote:
> Hi com.dsp-ers,
> 
> I know that there are some scenarios in which a line echo canceller
> should suspend the process of adapting its coefficients:
> 
> 1. Double talk (the most obvious case)
> 2. Narrow band reference signal detected
> 3. Numerical overload detected
> 
> I wanted to broadcast a call for ideas for detecting scenarios 2 and
> 3. For 2, I am currently thinking of putting in an 8 channel filter
> bank and using the relation between a short term estimate of power
> recorded at the output of the 8 channels to infer narrow band signals
> and freeze adaptation. For numerical overload, short term estimate of
> reference signal power is calculated and adaptation frozen in case a
> defined upper limit is crossed. So, what I would like to ask is:
> 
> 1. Are there any obvious deficiencies (that I am overlooking) in my
> approach.
> 2. Is there a better/standard way of doing either of the two tasks? (I
> could not find much in the several EC papers I went through, but these
> were mainly pretty old papers by pioneers of EC).
> 
> Any help is greatly appreaciated.

There is very little information about anything important in echo 
cancellation on the internet. There is a lot about the blatantly obvious 
basics. There is also a lot about achieving the fastest convergence - 
e.g numerous papers on FAP. That is seldom make or break for a design, 
though. When it comes to the subtler things, like you are looking at now 
with narrowband energy, there is very little info.

I don't think there is currently a "standard" way to look for narrowband 
energy in a canceller. An FFT might be one way. What happens if your 
spectral line signal is rich in harmonics, though? Looking for a peaky 
autocorrelation function certainly works, though it talks some compute. 
Low complexity things which directly look for narrowband energy, such as 
TKEO, might be more direct.

Steve

eternal_nan@yahoo.com wrote:

> Hi com.dsp-ers,
> 
> I know that there are some scenarios in which a line echo canceller
> should suspend the process of adapting its coefficients:
> 
> 1. Double talk (the most obvious case)

This is the most important and not so obvious case.

> 2. Narrow band reference signal detected

First option: you can trigger on the typical narrowband signals like 
DTMF, dialtone and other signaling tones.
Second option: a narrowband signal can be approximately detected by the 
energy threshold and the average peak to RMS ratio.

> 3. Numerical overload detected

This one is relatively simple, however it depends on the hardware.

> 
> I wanted to broadcast a call for ideas for detecting scenarios 2 and
> 3. For 2, I am currently thinking of putting in an 8 channel filter
> bank and using the relation between a short term estimate of power
> recorded at the output of the 8 channels to infer narrow band signals
> and freeze adaptation. For numerical overload, short term estimate of
> reference signal power is calculated and adaptation frozen in case a
> defined upper limit is crossed.  So, what I would like to ask is:
> 
> 1. Are there any obvious deficiencies (that I am overlooking) in my
> approach.

#2: It may work although it is complex and demands a lot of computation.
#3: Not a good idea, since you don't know what is the coupling between 
the transmit and receive path.

> 2. Is there a better/standard way of doing either of the two tasks?

Keep it simple. The main advantage of the SG LMS adaptation is its 
simplicity and low computation requirements.

> could not find much in the several EC papers I went through, but these
> were mainly pretty old papers by pioneers of EC).

I doubt if you can find any advanced information other then the common 
knowledge from the textbooks.



Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

Thanks for the advice guys.

On Mar 11, 12:42 am, Vladimir Vassilevsky <antispam_bo...@hotmail.com>
wrote:
> eternal_...@yahoo.com wrote:
> > Hi com.dsp-ers,
>
> > I know that there are some scenarios in which a line echo canceller
> > should suspend the process of adapting its coefficients:
>
> > 1. Double talk (the most obvious case)
>
> This is the most important and not so obvious case.
>
> > 2. Narrow band reference signal detected
>
> First option: you can trigger on the typical narrowband signals like
> DTMF, dialtone and other signaling tones.
> Second option: a narrowband signal can be approximately detected by the
> energy threshold and the average peak to RMS ratio.
>
> > 3. Numerical overload detected
>
> This one is relatively simple, however it depends on the hardware.
>
>
>
> > I wanted to broadcast a call for ideas for detecting scenarios 2 and
> > 3. For 2, I am currently thinking of putting in an 8 channel filter
> > bank and using the relation between a short term estimate of power
> > recorded at the output of the 8 channels to infer narrow band signals
> > and freeze adaptation. For numerical overload, short term estimate of
> > reference signal power is calculated and adaptation frozen in case a
> > defined upper limit is crossed.  So, what I would like to ask is:
>
> > 1. Are there any obvious deficiencies (that I am overlooking) in my
> > approach.
>
> #2: It may work although it is complex and demands a lot of computation.
> #3: Not a good idea, since you don't know what is the coupling between
> the transmit and receive path.
>
> > 2. Is there a better/standard way of doing either of the two tasks?
>
> Keep it simple. The main advantage of the SG LMS adaptation is its
> simplicity and low computation requirements.
>
> > could not find much in the several EC papers I went through, but these
> > were mainly pretty old papers by pioneers of EC).
>
> I doubt if you can find any advanced information other then the common
> knowledge from the textbooks.
>
> Vladimir Vassilevsky
>
> DSP and Mixed Signal Design Consultant
>
> http://www.abvolt.com