Speex dev - Jun 2009 - Speech switching in speakerphone?

>> I think the residual echo estimation is fairly reliable but I do not
know
>> how to use this to improve Pframe and in that way solve our main
problem
>> with the gain during near end talk.
>
> How can you tell that the residual echo estimation is reliable? in any
> case, I suspect that the whole Pframe idea might have to be revised
> (i.e. computing it completely differently or even not having a Pframe at
> all).
My impression when simulating in matlab is that the noise estimate which is
based on the residual echo is correlating well with the far-end activity.

>> Our main problem with the poor reliability during near end talk is what
we
>> need to find a solution to now. 
>
> OK, I'll have to think about it. At this point I'm not sure what to
do
> here. If you find anything, I'm definitely interested.
The one idea I have is to have a default gain2 of 0dB and only when the 
residual echo goes high we reduce the gain2. This could be implemented
similar to eq. (3) and (4) in this paper:

http://research.microsoft.com/pubs/69504/diegobenderskyhdres.pdf

The idea would be to use your estimation of the residual echo and noise
But to modify the calculation of the gain2. This would probably offer 
a more robust solution since we do not have to rely on the SNR. However 
we would lose the noise suppression functionality I guess (which probably 
is ok).

Are you familiar with any other RES-algorithm that might be more robust
to weak near-end signals?

Best regards
/Johan

> My impression when simulating in matlab is that the noise estimate which is
> based on the residual echo is correlating well with the far-end activity.
Not sure what you mean here.
> The idea would be to use your estimation of the residual echo and noise
> But to modify the calculation of the gain2. This would probably offer 
> a more robust solution since we do not have to rely on the SNR. However 
> we would lose the noise suppression functionality I guess (which probably 
> is ok).
Actually, if you remove the noise suppression, then it will sound
strange when you remove echo, because the noise will suddenly go away at
the same time. This is a bit annoying.
> Are you familiar with any other RES-algorithm that might be more robust
> to weak near-end signals?
Not really. This isn't something I have studied too much to be honest.

	Jean-Marc

> > My impression when simulating in matlab is that the noise estimate
which is
> > based on the residual echo is correlating well with the far-end
activity.
>
> Not sure what you mean here.
I mean that the echo_noise (based on residual echo) is showing a distinct
response
when fare end is talking. This in contrast to the SNR which is not always
showing
distinct response when near end is talking (due to the weak signal). 
> > The idea would be to use your estimation of the residual echo and
noise
> > But to modify the calculation of the gain2. This would probably offer 
> > a more robust solution since we do not have to rely on the SNR.
However
> > we would lose the noise suppression functionality I guess (which
probably
> > is ok).
>
> Actually, if you remove the noise suppression, then it will sound
> strange when you remove echo, because the noise will suddenly go away at
> the same time. This is a bit annoying.
Yes you are right about this but unfortunately we might be forced to accept this
or we have to insert synthetic noise during suppression. 
> > Are you familiar with any other RES-algorithm that might be more
robust
> > to weak near-end signals?
>
> Not really. This isn't something I have studied too much to be honest.
After some thinking I realized that the algorithm I was looking for actually
could
be realized with your preprocessor if I set both echo_suppress_active and
noise_suppress to 0 dB and echo_suppression to some appropriate damping.

I tried this out but it did not work initially. I tried to set all three
suppress
parameters to 0dB which should give a gain_floor of 0dB, but it did not.
I had to take a closer look on the fixed point version of the compute_gain_floor
function. It turned out that the addition of "1" you have in the
denominator
had to be removed in order to make it work as I expected it to do. I also
hade to change the gain_ratio multiplication command to MULT16_32_P15. I do 
not know if this is to consider as a bug or if I was pushing the function too
far.

The equation I use now looks like this:
gain_floor[i] =
MULT16_16_Q15(noise_gain,spx_sqrt(SHL32(EXTEND32(DIV32_16_Q15(PSHR32(noise[i],NOISE_SHIFT)
+ MULT16_32_P15(gain_ratio,echo[i]),(0+PSHR32(noise[i],NOISE_SHIFT) + echo[i])
)),15)));

Now the preprocessor behaves like the algorithm I was looking for!
 
Is there any reason you have the addition of 1 in the denominator?

Best Regards

Johan