Reply by maury December 11, 20082008-12-11
On Dec 11, 4:42&#4294967295;am, Lloydchristmas <olivier.hochreuti...@gmail.com>
wrote:
> On Dec 11, 2:11&#4294967295;am, ste...@coppice.org wrote: > > > > > > > On Dec 11, 4:17&#4294967295;am, maury <maury...@core.com> wrote: > > > > On Dec 9, 2:54&#4294967295;am, Lloydchristmas <olivier.hochreuti...@gmail.com> > > > wrote: > > > > > On Dec 8, 6:39&#4294967295;pm, ste...@coppice.org wrote: > > > > > > On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > > > > > wrote: > > > > > > > pal.debabrata123 wrote: > > > > > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > > > > > Why only 64 samples? > > > > > > To make sure it doesn't work too well? :-\ > > > > > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > > > > > linear equation? Input is known, output is known , only unknon is > > > > > > > coefficient? > > > > > > > It is certainly possible. > > > > > > > > Has anybody tried this ? > > > > > > > Kalman and Wiener tried this. > > > > > > I don't think Kalman's work has featured too much in echo > > > > > cancellation, but Wiener's certainly does. > > > > > > > > Is it too inefficient? > > > > > > > It works very well. > > > > > > A solution is only really inefficient if you have a simpler > > > > > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > > > > > they all have their roots in Weiner's work. Nothing fundamentally > > > > > better has come along.They take a lot of compute, so some brilliant > > > > > new solution would be most welcome. Don't hold your breath, though. > > > > > Actually, there is something new and it is working really well (and it > > > > is less complex than adaptive filtering) :http://lcavwww.epfl.ch/~cfaller/mypapers/faller03_AES114_PEC.pdf > > > > > Olivier Hochreutiner > > > > DSP/Embedded Software Engineer- Hide quoted text - > > > > > - Show quoted text - > > > > Oliver, > > > I suggest you look at Faller's paper a little more closely. &#4294967295;He is > > > using NLMS. &#4294967295;What he is describing is a sub-band NLMS algorithm. &#4294967295;This > > > was originally described by Widrow in the 1960's. &#4294967295;Several techniques > > > have been used to decompose the input signal, including FFT DCT, > > > wavelats, FSF, etc. &#4294967295;It has been called by different names such as sub- > > > band decomposition, sub-band LMS, and frequency-domain LMS to name a > > > few. > > > > Maurice Givens > > > They mention the use of NLMS, RLS, FAP or other adaption schemes, so > > its not just an NLMS paper, but its most certainly a conventional > > adaptive filter one. > > > They use the word perceptual like they have added a new twist to sub- > > band EC, but I couldn't really see anything novel. Maybe I missed an > > interesting new element. > > > Regards, > > Steve > > Actually it is not a conventional linear adaptive filtering technique. > They estimate the echo response in term of its spectral envelope and > then perform spectral subtraction (in a similar way than noise > suppression). The 'psychoacoustic' element is that they divide the > frequency spectrum in a 'psychoacoustic-friendly' way. > They indeed use NLMS or other adaptive algorithms for the spectral > estimation because it's an effective adaptive altorithm, but it's not > an classic adaptive filter. > There are more details in this paper:http://infoscience.epfl.ch/record/54882/files/ > > This psychoacoustic echo suppressor performs really better where the > acoustic path is not known at all in the first place (e.g. VoIP > software on a PC), or when the echo path has non-linear components > (e.g. saturation at the mic or speaker). > > Olivier- Hide quoted text - > > - Show quoted text -
Oliver, This really is a conventional technique. Yes, Faller shapes the sub- band by making the bins to match what he neesds, but this is not new. Yes, he uses spectral subtraction to perform the suppression, but this is not new either. The only novel thing I see in the paper is his particular shaping of the sub-band frequencies. If you can find a copy of it, look at the Proceedings of the Comp.Dsp Conference that was held some years ago. In there is an explantion of the spectral subtraction technique, and how it is used. The advantage of the technique is that it is not a parametric technique, but it is in no way a new or novel technique. It was developed expressly because the parameters of the unknown noise was not known and could not be determined effectively. This technique has been used in hearing aids for decades. It has been used in the telecommunications industry as well. It appears that you are constraining your idea of a "classic" adaptive filter as one that is LMS based. Maurice Givens
Reply by Lloydchristmas December 11, 20082008-12-11
On Dec 11, 2:11&#4294967295;am, ste...@coppice.org wrote:
> On Dec 11, 4:17&#4294967295;am, maury <maury...@core.com> wrote: > > > > > On Dec 9, 2:54&#4294967295;am, Lloydchristmas <olivier.hochreuti...@gmail.com> > > wrote: > > > > On Dec 8, 6:39&#4294967295;pm, ste...@coppice.org wrote: > > > > > On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > > > > wrote: > > > > > > pal.debabrata123 wrote: > > > > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > > > > Why only 64 samples? > > > > > To make sure it doesn't work too well? :-\ > > > > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > > > > linear equation? Input is known, output is known , only unknon is > > > > > > coefficient? > > > > > > It is certainly possible. > > > > > > > Has anybody tried this ? > > > > > > Kalman and Wiener tried this. > > > > > I don't think Kalman's work has featured too much in echo > > > > cancellation, but Wiener's certainly does. > > > > > > > Is it too inefficient? > > > > > > It works very well. > > > > > A solution is only really inefficient if you have a simpler > > > > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > > > > they all have their roots in Weiner's work. Nothing fundamentally > > > > better has come along.They take a lot of compute, so some brilliant > > > > new solution would be most welcome. Don't hold your breath, though. > > > > Actually, there is something new and it is working really well (and it > > > is less complex than adaptive filtering) :http://lcavwww.epfl.ch/~cfaller/mypapers/faller03_AES114_PEC.pdf > > > > Olivier Hochreutiner > > > DSP/Embedded Software Engineer- Hide quoted text - > > > > - Show quoted text - > > > Oliver, > > I suggest you look at Faller's paper a little more closely. &#4294967295;He is > > using NLMS. &#4294967295;What he is describing is a sub-band NLMS algorithm. &#4294967295;This > > was originally described by Widrow in the 1960's. &#4294967295;Several techniques > > have been used to decompose the input signal, including FFT DCT, > > wavelats, FSF, etc. &#4294967295;It has been called by different names such as sub- > > band decomposition, sub-band LMS, and frequency-domain LMS to name a > > few. > > > Maurice Givens > > They mention the use of NLMS, RLS, FAP or other adaption schemes, so > its not just an NLMS paper, but its most certainly a conventional > adaptive filter one. > > They use the word perceptual like they have added a new twist to sub- > band EC, but I couldn't really see anything novel. Maybe I missed an > interesting new element. > > Regards, > Steve
Actually it is not a conventional linear adaptive filtering technique. They estimate the echo response in term of its spectral envelope and then perform spectral subtraction (in a similar way than noise suppression). The 'psychoacoustic' element is that they divide the frequency spectrum in a 'psychoacoustic-friendly' way. They indeed use NLMS or other adaptive algorithms for the spectral estimation because it's an effective adaptive altorithm, but it's not an classic adaptive filter. There are more details in this paper: http://infoscience.epfl.ch/record/54882/files/ This psychoacoustic echo suppressor performs really better where the acoustic path is not known at all in the first place (e.g. VoIP software on a PC), or when the echo path has non-linear components (e.g. saturation at the mic or speaker). Olivier
Reply by December 10, 20082008-12-10
On Dec 11, 4:17&#4294967295;am, maury <maury...@core.com> wrote:
> On Dec 9, 2:54&#4294967295;am, Lloydchristmas <olivier.hochreuti...@gmail.com> > wrote: > > > > > On Dec 8, 6:39&#4294967295;pm, ste...@coppice.org wrote: > > > > On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > > > wrote: > > > > > pal.debabrata123 wrote: > > > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > > > Why only 64 samples? > > > > To make sure it doesn't work too well? :-\ > > > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > > > linear equation? Input is known, output is known , only unknon is > > > > > coefficient? > > > > > It is certainly possible. > > > > > > Has anybody tried this ? > > > > > Kalman and Wiener tried this. > > > > I don't think Kalman's work has featured too much in echo > > > cancellation, but Wiener's certainly does. > > > > > > Is it too inefficient? > > > > > It works very well. > > > > A solution is only really inefficient if you have a simpler > > > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > > > they all have their roots in Weiner's work. Nothing fundamentally > > > better has come along.They take a lot of compute, so some brilliant > > > new solution would be most welcome. Don't hold your breath, though. > > > Actually, there is something new and it is working really well (and it > > is less complex than adaptive filtering) :http://lcavwww.epfl.ch/~cfaller/mypapers/faller03_AES114_PEC.pdf > > > Olivier Hochreutiner > > DSP/Embedded Software Engineer- Hide quoted text - > > > - Show quoted text - > > Oliver, > I suggest you look at Faller's paper a little more closely. &#4294967295;He is > using NLMS. &#4294967295;What he is describing is a sub-band NLMS algorithm. &#4294967295;This > was originally described by Widrow in the 1960's. &#4294967295;Several techniques > have been used to decompose the input signal, including FFT DCT, > wavelats, FSF, etc. &#4294967295;It has been called by different names such as sub- > band decomposition, sub-band LMS, and frequency-domain LMS to name a > few. > > Maurice Givens
They mention the use of NLMS, RLS, FAP or other adaption schemes, so its not just an NLMS paper, but its most certainly a conventional adaptive filter one. They use the word perceptual like they have added a new twist to sub- band EC, but I couldn't really see anything novel. Maybe I missed an interesting new element. Regards, Steve
Reply by CW December 10, 20082008-12-10
On Dec 8, 12:39&#4294967295;pm, ste...@coppice.org wrote:
> On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > wrote: > > > pal.debabrata123 wrote: > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > Why only 64 samples? > > To make sure it doesn't work too well? :-\ > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > linear equation? Input is known, output is known , only unknon is > > > coefficient? > > > It is certainly possible. > > > > Has anybody tried this ? > > > Kalman and Wiener tried this. > > I don't think Kalman's work has featured too much in echo > cancellation, but Wiener's certainly does. > > > > > > Is it too inefficient? > > > It works very well. > > A solution is only really inefficient if you have a simpler > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > they all have their roots in Weiner's work. Nothing fundamentally > better has come along. They take a lot of compute, so some brilliant > new solution would be most welcome. Don't hold your breath, though. > > Regards, > Steve
Solution to the Normal equations. http://www.elec.york.ac.uk/comms/pdfs/20060427123418.pdf CW
Reply by maury December 10, 20082008-12-10
On Dec 9, 2:54&#4294967295;am, Lloydchristmas <olivier.hochreuti...@gmail.com>
wrote:
> On Dec 8, 6:39&#4294967295;pm, ste...@coppice.org wrote: > > > > > > > On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > > wrote: > > > > pal.debabrata123 wrote: > > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > > Why only 64 samples? > > > To make sure it doesn't work too well? :-\ > > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > > linear equation? Input is known, output is known , only unknon is > > > > coefficient? > > > > It is certainly possible. > > > > > Has anybody tried this ? > > > > Kalman and Wiener tried this. > > > I don't think Kalman's work has featured too much in echo > > cancellation, but Wiener's certainly does. > > > > > Is it too inefficient? > > > > It works very well. > > > A solution is only really inefficient if you have a simpler > > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > > they all have their roots in Weiner's work. Nothing fundamentally > > better has come along.They take a lot of compute, so some brilliant > > new solution would be most welcome. Don't hold your breath, though. > > Actually, there is something new and it is working really well (and it > is less complex than adaptive filtering) :http://lcavwww.epfl.ch/~cfaller/mypapers/faller03_AES114_PEC.pdf > > Olivier Hochreutiner > DSP/Embedded Software Engineer- Hide quoted text - > > - Show quoted text -
Oliver, I suggest you look at Faller's paper a little more closely. He is using NLMS. What he is describing is a sub-band NLMS algorithm. This was originally described by Widrow in the 1960's. Several techniques have been used to decompose the input signal, including FFT DCT, wavelats, FSF, etc. It has been called by different names such as sub- band decomposition, sub-band LMS, and frequency-domain LMS to name a few. Maurice Givens
Reply by Lloydchristmas December 9, 20082008-12-09
On Dec 8, 6:39&#4294967295;pm, ste...@coppice.org wrote:
> On Dec 9, 12:20&#4294967295;am, Vladimir Vassilevsky <antispam_bo...@hotmail.com> > wrote: > > > pal.debabrata123 wrote: > > > Ok , assume echo tail length is less that or equal to 64 sample. > > > Why only 64 samples? > > To make sure it doesn't work too well? :-\ > > > > Is it possible to get coefficient of the adaptive filter by solving 64 > > > linear equation? Input is known, output is known , only unknon is > > > coefficient? > > > It is certainly possible. > > > > Has anybody tried this ? > > > Kalman and Wiener tried this. > > I don't think Kalman's work has featured too much in echo > cancellation, but Wiener's certainly does. > > > > > > Is it too inefficient? > > > It works very well. > > A solution is only really inefficient if you have a simpler > alternative. Various forms of NLMS, PNLMS, FAP are used today, but > they all have their roots in Weiner's work. Nothing fundamentally > better has come along.They take a lot of compute, so some brilliant > new solution would be most welcome. Don't hold your breath, though.
Actually, there is something new and it is working really well (and it is less complex than adaptive filtering) : http://lcavwww.epfl.ch/~cfaller/mypapers/faller03_AES114_PEC.pdf Olivier Hochreutiner DSP/Embedded Software Engineer
Reply by mmoctar December 9, 20082008-12-09
>Ok , assume echo tail length is less that or equal to 64 sample. > >Is it possible to get coefficient of the adaptive filter by solving 64 >linear equation? Input is known, output is known , only unknon is >coefficient? > >Has anybody tried this ? Is it too inefficient? >
I think that it's depend on the length of your input signal,if it's enougth to converge you can get your filter coef. If you are not in a pratical way, you just need the filter coefs, you can you use the Wiener filter.
Reply by Vladimir Vassilevsky December 8, 20082008-12-08

maury wrote:


> General question for the group. Let the unknown system be an unstable > system. Can the LMS algorithm still estimate the coefficients of the > impulse response?
Yes. As long as the system diverges slower then the LMS converges. The LMS can catch up even if the echo canceller came into the oscillation. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
Reply by December 8, 20082008-12-08
On Dec 8, 10:46&#4294967295;am, "pal.debabrata123" <pal.debabrata...@gmail.com>
wrote:
> Ok , assume echo tail length is less that or equal to 64 sample. > > Is it possible to get coefficient of the adaptive filter by solving 64 > linear equation? Input is known, output is known , only unknon is > coefficient? > > Has anybody tried this ? Is it too inefficient?
This can work but there is a potential pitfall. If your data has a phase ambiguity between the signal's sample clock and its modulation, then the matrix method (formed by setting up a system of the linear equations) will be "too exact" in that it tries to correct the phase issues as well. So if the solution resulting from the matrix method is applied to another block of data with a difference phase offset, the results probably won't be good. I've seen in this case where the matrix solution introduces high frequency artifacts into the "corrected" signal. On the other hand a block of data that has a reference signal in part of it, then the matrix solution can work wonderfully well for the rest of the data block providied the sample clock has a high temporal coherence. An LMS approach with reasonably slow adaptation can average out the phase issues for nonphase locked situations. Clay
Reply by maury December 8, 20082008-12-08
On Dec 8, 9:46&#4294967295;am, "pal.debabrata123" <pal.debabrata...@gmail.com>
wrote:
> Ok , assume echo tail length is less that or equal to 64 sample. > > Is it possible to get coefficient of the adaptive filter by solving 64 > linear equation? Input is known, output is known , only unknon is > coefficient? > > Has anybody tried this ? Is it too inefficient?
Most echo cancellers use some form of the LMS algorithm. And, for stability with varying input levels, the NLMS is by far the algorithm of choice. Look at what the LMS is trying to do. It is trying to construct the Wiener filter R^-1(x,x)r(x,d). R is the autocorrelation matrix of the input, and r is the cross corrleation of the input and desired. So, yes, 64 equations are solved to give a 64-dimensional vector. HOWEVER, at what price? Are you going to have an accurate estimate of the corellations? What happens if the autocorrelation matrix is not a "nice" matrix? What happens if you can't find an inverse? Yes, you can do it, but is it worth it? General question for the group. Let the unknown system be an unstable system. Can the LMS algorithm still estimate the coefficients of the impulse response? Maurice Givens