Regarding double talk detection in Acoustic Echo Cancellation

Dear Sir,

              I am working on Acoustic Echo Cancellation using LMS
algorithm in time domain.I have written the code for this.It is working
fine that its output signal has only the near end signal with a very
very small hiss sound in the situation when both farend and near end
signals are talked simultaneously. As large eigen value spread  and
double talk situation are two major drawbacks of LMS algorithm for
Acoustic Echo Cancellation(AEC), for large eigen value spread I have
introduced Elliptic IIR filter and Polyphase IIR filter and I have
integrated them in the code of AEC  separately and I observed its
outputs separatelY.

             But I dont see much difference in the outputs of basic AEC
using LMS and the AEC with subband approach.They are almost similar
with some hiss sound in the double talk situation.So now I am seeing
Double Talk Detection(DTD).

             We know that double talk situation is nothing but the
situation when both farend and near end speach signals occur at a time,
during which there will be poor AEC performance.To avoid this when DT
occurs we need to freeze the updation of weights in AEC.

             But the problem is, how to detect double talk situation?It
is mentioned in some documents that basing on the signal energies we
will detect the double talk situation by setting some threshhold.But it
is not mentioned clearly that how to take this threshhold value.

           If anyone have idea regarding this please give some valuable
suggestions.

                                                Thanks in advance.

Regards,
Aparna Ram.K.

Aparna Ram wrote:


>              But the problem is, how to detect double talk situation?It
> is mentioned in some documents that basing on the signal energies we
> will detect the double talk situation by setting some threshhold.

It is dead simple.
The transmitted power is known, the received power is known, the 
coupling between the speaker and the mike is also known approximately. 
Therefore you can set a threshold based on transmit/receive power ratio. 
If the ratio is below the threshold, that means that the received signal 
is dominated by the sound from the ambience.

Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

Aparna Ram wrote:
> Dear Sir,
> 
>               I am working on Acoustic Echo Cancellation using LMS
> algorithm in time domain.I have written the code for this.It is working
> fine that its output signal has only the near end signal with a very
> very small hiss sound in the situation when both farend and near end
> signals are talked simultaneously. As large eigen value spread  and
> double talk situation are two major drawbacks of LMS algorithm for
> Acoustic Echo Cancellation(AEC), for large eigen value spread I have
> introduced Elliptic IIR filter and Polyphase IIR filter and I have
> integrated them in the code of AEC  separately and I observed its
> outputs separatelY.
> 
>              But I dont see much difference in the outputs of basic AEC
> using LMS and the AEC with subband approach.They are almost similar
> with some hiss sound in the double talk situation.So now I am seeing
> Double Talk Detection(DTD).
> 
>              We know that double talk situation is nothing but the
> situation when both farend and near end speach signals occur at a time,
> during which there will be poor AEC performance.To avoid this when DT
> occurs we need to freeze the updation of weights in AEC.
> 
>              But the problem is, how to detect double talk situation?It
> is mentioned in some documents that basing on the signal energies we
> will detect the double talk situation by setting some threshhold.But it
> is not mentioned clearly that how to take this threshhold value.

The commonest approach is something called the Geigel algorithm. Very 
simple, very crude, but rather flaky. Its simplicity means its a good 
place to start, but it probably won't build you a world class echo 
canceller. :-\

Basically you use a decision like "if the short term receive power is 
greater than half the short term transmit power it can't all be echo, so 
we have double talk". There are various temporal assumptions in that, 
and it assumes the far end talker will be fairly loud. Its is, thus, a 
far from perfect decision criterion.


>            If anyone have idea regarding this please give some valuable
> suggestions.

Steve