Speex dev - Feb 2011 - About Sampling Rate Correction in acoustic echo

Thank you, Andreas Engel.
I downloaded the white paper of the Fraunhofer Acoustic Echo Control.
http://www.iis.fraunhofer.de/bf/amm/download/whitepapers/Acoustic_Echo_Control-wp.pdf
It said 
> "In the Fraunhofer Acoustic Echo Control, the frequency spectrum of
the microphone signal is
> modified so that the undesired echo components are removed from the signal
transmitted to
> the far-end. The general approach is illustrated in figure 2."
This modification is a gain multiplication in each frequency instead of
frequency domain
subtraction in common frequency domain echo cancellation. See
...
> As can be seen, both the loudspeaker and microphone signals are first
transformed into the
> frequency domain by a spectral transform (ST). Based on these input
signals, the control unit of
> the Acoustic Echo Control determines an optimum gain factor for each
individual frequency
> band separately. These frequency dependent gain factors are also referred
to as echo
> attenuation filters. Obviously, the gain factors are selected to be close
to zero in circumstances
> where strong echo components need to be removed from the microphone signal.
On the other
> hand, in the event of near-end speech only, it is set to one in order to
leave the desired speech
> signal unchanged. After applying this echo attenuation filter to the
spectral representation of
> the microphone signal, the echo-free signal is transformed back to the time
domain by a
> corresponding inverse spectral transform (IST). In typical application
scenarios, robust
> attenuation of the echo by 60 dB can be expected and achieved reliably.
So it likes some kinds of a Noice Reducer. It's kernel is not common time-
or frequency-domain
adaptive filter. It is not sensitive to tiny frequency shift but the phase of
the echo is lost.
So I doubt the quality of voice after processing especially in double
talk.
> The calculation of the optimum gain factor is based on an estimate of the
power spectrum of
> the echo signal captured by the microphone. The power spectrum of the echo
is determined by
> applying an adaptive estimate of the acoustic echo path to the known power
spectrum of the
> loudspeaker signal
There is still a question. Which algorithm is this adaptive echo power spectrum
estimation based on?
Is this algorithm not sensitive to frequency difference?
Sincerely
Maoquan
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/speex-dev/attachments/20110210/a558c816/attachment.htm

I can only evaluate this with my subjective point of view. I had a 
special test scenario doing chat with cheap webcam microphones and 
loudspeakers. Fraunhofers solution was the only one that could eliminate 
the echo. In double talk the quality gets lower but is still very good. 
You might want to ask Fraunhofer for a demo version to test for yourself.
I have no details on the algorithms being used, I only know that it is 
patented. Since it is a commercial product I doubt they are willing to 
share any more details than they already do in that paper. But the 
algorithm is obviously very insensitive to frequency
drift.
>
> Thank you, Andreas Engel.
>
> I downloaded the white paper of the Fraunhofer Acoustic Echo Control.
>
>
http://www.iis.fraunhofer.de/bf/amm/download/whitepapers/Acoustic_Echo_Control-wp.pdf
>
> It said
>
> > "In the Fraunhofer Acoustic Echo Control, the frequency spectrum
of
> the microphone signal is
> > modified so that the undesired echo components are removed from the 
> signal transmitted to
> > the far-end. The general approach is illustrated in figure 2."
>
> This modification is a gain multiplication in each frequency instead 
> of frequency domain
> subtraction in common frequency domain echo cancellation. See ...
>
> > As can be seen, both the loudspeaker and microphone signals are 
> first transformed into the
> > frequency domain by a spectral transform (ST). Based on these input 
> signals, the control unit of
> > the Acoustic Echo Control determines an optimum gain factor for each 
> individual frequency
> > band separately. These frequency dependent gain factors are also 
> referred to as echo
> > attenuation filters. Obviously, the gain factors are selected to be 
> close to zero in circumstances
> > where strong echo components need to be removed from the microphone 
> signal. On the other
> > hand, in the event of near-end speech only, it is set to one in 
> order to leave the desired speech
> > signal unchanged. After applying this echo attenuation filter to the 
> spectral representation of
> > the microphone signal, the echo-free signal is transformed back to 
> the time domain by a
> > corresponding inverse spectral transform (IST). In typical 
> application scenarios, robust
> > attenuation of the echo by 60 dB can be expected and achieved
reliably.
>
> So it likes some kinds of a Noice Reducer. It's kernel is not common 
> time- or frequency-domain
> adaptive filter. It is not sensitive to tiny frequency shift but the 
> phase of the echo is lost.
> So I doubt the quality of voice after processing especially in double 
> talk.
>
> > The calculation of the optimum gain factor is based on an estimate 
> of the power spectrum of
> > the echo signal captured by the microphone. The power spectrum of 
> the echo is determined by
> > applying an adaptive estimate of the acoustic echo path to the known 
> power spectrum of the
> > loudspeaker signal
>
> There is still a question. Which algorithm is this adaptive echo power 
> spectrum estimation based on?
> Is this algorithm not sensitive to frequency difference?
>
> Sincerely
> Maoquan
>
>
> _______________________________________________
> Speex-dev mailing list
> Speex-dev at xiph.org
> http://lists.xiph.org/mailman/listinfo/speex-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/speex-dev/attachments/20110210/4e1db622/attachment.htm

On 02/10/2011 06:21 PM, Andreas Engel wrote:
> I can only evaluate this with my subjective point of view. I had a 
> special test scenario doing chat with cheap webcam microphones and 
> loudspeakers. Fraunhofers solution was the only one that could 
> eliminate the echo. In double talk the quality gets lower but is still 
> very good. You might want to ask Fraunhofer for a demo version to test 
> for yourself.
> I have no details on the algorithms being used, I only know that it is 
> patented. Since it is a commercial product I doubt they are willing to 
> share any more details than they already do in that paper. But the 
> algorithm is obviously very insensitive to frequency drift.
If they charge a lot, the software will almost certainly make use of 
some interesting patented techniques. Echo cancellation is a patent 
minefield. A patent search with "Fraunhofer" might drag up something 
interesting.

Steve