Forums

Echo canceller

Started by Luiz Carlos October 4, 2004
White noise is a good, if not the best, training signal for the echo
canceller.
White noise has all the frequencies from 0 to fs/2.
In telephony the range of usefull spectrum is roughly from 200Hz to
3400Hz (G712). So the output white noise will come back (echo) as a
colored noise (band filtered  white noise).
So the echo canceller has to imitate this filtering.

What is better for training the echo canceller ?
a) Use a real white noise.
b) Use a filtered white noise, to free the echo canceller from doing
this filtering.

Luiz Carlos
oen_no_spam@yahoo.com.br (Luiz Carlos) wrote in message news:<3fd8f66b.0410040327.44845258@posting.google.com>...
> White noise is a good, if not the best, training signal for the echo > canceller. > White noise has all the frequencies from 0 to fs/2. > In telephony the range of usefull spectrum is roughly from 200Hz to > 3400Hz (G712). So the output white noise will come back (echo) as a > colored noise (band filtered white noise). > So the echo canceller has to imitate this filtering. > > What is better for training the echo canceller ? > a) Use a real white noise. > b) Use a filtered white noise, to free the echo canceller from doing > this filtering. > > Luiz Carlos
Luiz, That depends on what you are trying to accomplish. If convergence speed is what you seek, then a single frequency sinusoid is the best. You can typically converge in about 2 or 3 samples. If you are trying to converge to 'all' signals, then wideband white noise is required. If you are trying to converge in a manner that would be similar to speech, then it is argued that the composite-source signal is best (ITU-G.168) Maurice Givens
> Luiz, > That depends on what you are trying to accomplish. If convergence > speed is what you seek, then a single frequency sinusoid is the best. > You can typically converge in about 2 or 3 samples. If you are trying > to converge to 'all' signals, then wideband white noise is required. > If you are trying to converge in a manner that would be similar to > speech, then it is argued that the composite-source signal is best > (ITU-G.168) > > > Maurice Givens
Maurice, What I'm trying to undertand is if I am making thigs difficult for the estimator when I use white noise, because the codec filters out low and high frequencies, . If this natural (from the codecs) filter has an equivalent FIR lenght higher than the lenght of the echo canceller, it may complicate the work of the estimator. Which algorithm converges in 3 samples using a single sinusoid? Luiz Carlos.
oen_no_spam@yahoo.com.br (Luiz Carlos) wrote in message news:<3fd8f66b.0410050316.ef703cd@posting.google.com>...
> > Luiz, > > That depends on what you are trying to accomplish. If convergence > > speed is what you seek, then a single frequency sinusoid is the best. > > You can typically converge in about 2 or 3 samples. If you are trying > > to converge to 'all' signals, then wideband white noise is required. > > If you are trying to converge in a manner that would be similar to > > speech, then it is argued that the composite-source signal is best > > (ITU-G.168) > > > > > > Maurice Givens > > Maurice, > > What I'm trying to undertand is if I am making thigs difficult for the > estimator when I use white noise, because the codec filters out low > and high frequencies, . > If this natural (from the codecs) filter has an equivalent FIR lenght > higher than the lenght of the echo canceller, it may complicate the > work of the estimator. > > Which algorithm converges in 3 samples using a single sinusoid? > > Luiz Carlos.
I have never had a problem with white noise using A- or mu-law codecs. Other low-bit rate coders are a different question. The source you use is determined by your implementation. If you are using with voice, then arguably the CCS is best. If you are trying to characterize the canceller in general, then white noise. If you want to show a fast convergnce time for the spec sheet, use a sinusoid. As far as extremely quick convergence, make a normalized LMS algorithm model with a gain of 1 and feed it a single frequency sinusoid.