Forums

Acoustic feedback cancellation

Started by Roman Rumian October 1, 2007
Hello friends,

We are working on an Acoustic Feedback Cancellation system. The system 
must cancel the feedback tone before this (usually pure tone) becomes 
dominant or even audible.
FFT methods (for the howling frequency detection and notch adaptation) 
developed by my friend seems to be too slow and acoustic howling becomes 
audible for about one second. The method could be speed up if the fast 
pure tone extraction algorithm would be added before. Recently we are 
looking for a good algorithm for a pure tone (also weak) 
detection/extraction from music/speech signals. The LMS FIR methods 
(delayed LMS, filtered-x LMS) seems to be slow in adaptation or even 
unstable (fixed point application).
The LMS (e. g. sign-sign LMS 
http://ieeexplore.ieee.org/iel5/9008/28596/01279409.pdf ) Adaptive 
(single pole IIR) Notch Filter seems to be a good alternative, but could 
adapt to the non howling tone and degrade sound quality.
Could you help us to find the best adaptive pure tone extraction 
algorithm ?

Regards

Roman Rumian
"Roman Rumian" <usun_torumian@agh.edu.pl> wrote in message
news:fdqn63$jvk$1@news.agh.edu.pl...
> Hello friends, > > We are working on an Acoustic Feedback Cancellation system. The system > must cancel the feedback tone before this (usually pure tone) becomes > dominant or even audible. > FFT methods (for the howling frequency detection and notch adaptation) > developed by my friend seems to be too slow and acoustic howling becomes > audible for about one second. The method could be speed up if the fast > pure tone extraction algorithm would be added before. Recently we are > looking for a good algorithm for a pure tone (also weak) > detection/extraction from music/speech signals. The LMS FIR methods > (delayed LMS, filtered-x LMS) seems to be slow in adaptation or even > unstable (fixed point application). > The LMS (e. g. sign-sign LMS > http://ieeexplore.ieee.org/iel5/9008/28596/01279409.pdf ) Adaptive > (single pole IIR) Notch Filter seems to be a good alternative, but could > adapt to the non howling tone and degrade sound quality. > Could you help us to find the best adaptive pure tone extraction > algorithm ?
1. The LMS approach is obviously superior to adaptive notches. An adaptive notch is a simple fix for the problem; the LMS filter takes care of the cause. So, the LMS is the way to do it you can afford the computing cost. 2. It doesn't matter if you use a time domain or a frq domain approach. The very good way to detect "howling" is by the autocorrelation; however it takes a lot of computation. 3. You shouldn't expect the miraculous results. EC can buy the additional ~20dB of attenuation, however it would be difficult to get more than that in a practical situation (as opposed to the lab setup in the ideal conditions). 4. For the initial convergence of the EC filter, you can break the feedback path. For example, you can run some sort of prompt signal before opening the audio path, and train the LMS on that prompt. There is a lot of things that you can do depending on the application. Vladimir Vassilevsky DSP and Mixed Signal Consultant www.abvolt.com
Hello Vladimir,

Vladimir Vassilevsky napisa&#2013266099;(a):
(...)
> 1. The LMS approach is obviously superior to adaptive notches. An adaptive > notch is a simple fix for the problem; the LMS filter takes care of the > cause. So, the LMS is the way to do it you can afford the computing cost. > > 2. It doesn't matter if you use a time domain or a frq domain approach. The > very good way to detect "howling" is by the autocorrelation; however it > takes a lot of computation. > > 3. You shouldn't expect the miraculous results. EC can buy the additional > ~20dB of attenuation, however it would be difficult to get more than that > in a practical situation (as opposed to the lab setup in the ideal > conditions). > > 4. For the initial convergence of the EC filter, you can break the feedback > path. For example, you can run some sort of prompt signal before opening the > audio path, and train the LMS on that prompt. > > There is a lot of things that you can do depending on the application.
many thanks for your answer ! :-) This question was actually set by my room mate, and, of couse, he is finishing LMS approach. Kind regards Roman Rumian
Roman Rumian wrote:
> We are working on an Acoustic Feedback Cancellation system. The system > must cancel the feedback tone before this (usually pure tone) becomes > dominant or even audible.
> Could you help us to find the best adaptive pure tone extraction > algorithm ?
Are you processing music or a speaker's microphone? In the latter case there is a ditry but very effective trick by shifting the frequency of the signal a little bit. This will reduce the critical feedback level by effectively smoothing the peaks of the transition function. The classical approach is to do a SSB modulation and demodulate with a slightly detuned frequency. Marcel