Prev: How to spec jitter for a digital NCO
Next: Literature/Link references?
From: srikanth.konj on 1 Jun 2010 06:24
Hi All,

my problem statement is as follows

Prob:
=====
a feedback signal which is a form of sinusoid or mix of sinusoids is
getting mixed to the speech signal. So i would like to remove the unwanted
sinusoidal signals. (here noise is sinusoid and not white noise)

Experiment:
===========
Trying to use LMS to remove the sunusoid. My understanding is that
LMS will initally try to autocorrelate the input signal to itself and the
coefficients will adapt as a bandpass filter with the center as the
sinusoidal frequency. So the error here is the speech signal after
subtraction with the desired signal.

Question:
=========
1. My intention is to increase the error signal (speech) so how to select
the step size (convergence index)
2. most of the theories try to reduce the mean square error and decide on
the step sizes. Will the same work here
3. How much delay is necessary to decrease the correlation of the speech
signal at the same time maintain the correlation of the inherent sinusoid
signal (noise here)



From: Vladimir Vassilevsky on 1 Jun 2010 10:59

1. What you are asking for is called "howling suppression".

2. LMS won't work. Because you don't have clean reference to optimize
with respect to.

3. What you need is adaptive notch filter. There are tons of boring
books and articles to read.

4. This is complicated area and very commertial also. Amateur approach
is not feasible. Don't expect a solution for free.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com




srikanth.konj wrote:
> Hi All,
>
> my problem statement is as follows
>
> Prob:
> =====
> a feedback signal which is a form of sinusoid or mix of sinusoids is
> getting mixed to the speech signal. So i would like to remove the unwanted
> sinusoidal signals. (here noise is sinusoid and not white noise)
>
> Experiment:
> ===========
> Trying to use LMS to remove the sunusoid. My understanding is that
> LMS will initally try to autocorrelate the input signal to itself and the
> coefficients will adapt as a bandpass filter with the center as the
> sinusoidal frequency. So the error here is the speech signal after
> subtraction with the desired signal.
>
> Question:
> =========
> 1. My intention is to increase the error signal (speech) so how to select
> the step size (convergence index)
> 2. most of the theories try to reduce the mean square error and decide on
> the step sizes. Will the same work here
> 3. How much delay is necessary to decrease the correlation of the speech
> signal at the same time maintain the correlation of the inherent sinusoid
> signal (noise here)
>
>
>
From: maury on 2 Jun 2010 12:47
On Jun 1, 5:24 am, "srikanth.konj"
<srikanth.konj(a)n_o_s_p_a_m.gmail.com> wrote:
> Hi All,
>
>   my problem statement is as follows
>
> Prob:
> =====
>   a feedback signal which is a form of sinusoid or mix of sinusoids is
> getting mixed to the speech signal. So i would like to remove the unwanted
> sinusoidal signals. (here noise is sinusoid and not white noise)
>
> Experiment:
> ===========
> Trying to use LMS to remove the sunusoid. My understanding is that
> LMS will initally try to autocorrelate the input signal to itself and the
> coefficients will adapt as a bandpass filter with the center as the
> sinusoidal frequency. So the error here is the speech signal after
> subtraction with the desired signal.
>
> Question:
> =========
> 1. My intention is to increase the error signal (speech) so how to select
> the step size (convergence index)
> 2. most of the theories try to reduce the mean square error and decide on
> the step sizes. Will the same work here
> 3. How much delay is necessary to decrease the correlation of the speech
> signal at the same time maintain the correlation of the inherent sinusoid
> signal (noise here)

Since this is an experiment, do it as an experiment.
I'll give you a hint, look up adaptive line enhancer (ALE). It is a
LMS-based adaptive notch filter that makes its own reference. Make a
simulation of the ALE, then run it varying the update gain and see the
results as a function of the update gain. Then decide if LMS-based
filters work in this application. If you decide they do, then what do
you need to do to optimize the results? What are the pitfalls? What
are the advantages? What are the disadvantages? Under what
circumstances would you use it? When would you not use it?

Maurice Givens
From: Vladimir Vassilevsky on 2 Jun 2010 13:17


maury wrote:

> On Jun 1, 5:24 am, "srikanth.konj"
> <srikanth.konj(a)n_o_s_p_a_m.gmail.com> wrote:
>
>> a feedback signal which is a form of sinusoid or mix of sinusoids is
>> getting mixed to the speech signal. So i would like to remove the unwanted
>> sinusoidal signals. (here noise is sinusoid and not white noise)


> I'll give you a hint, look up adaptive line enhancer (ALE). It is a
> LMS-based adaptive notch filter that makes its own reference. Make a
> simulation of the ALE, then run it varying the update gain and see the
> results as a function of the update gain. Then decide if LMS-based
> filters work in this application.

One approach to make the LMS work in this application is to add some
artificial components into the signal. Those components could be used as
a reference for system identification. The trick is how to make it
efficient without introducing the significant artifacts.

Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com
From: maury on 2 Jun 2010 17:14
On Jun 2, 12:17 pm, Vladimir Vassilevsky <nos...(a)nowhere.com> wrote:
> maury wrote:
> > On Jun 1, 5:24 am, "srikanth.konj"
> > <srikanth.konj(a)n_o_s_p_a_m.gmail.com> wrote:
>
> >> a feedback signal which is a form of sinusoid or mix of sinusoids is
> >> getting mixed to the speech signal. So i would like to remove the unwanted
> >> sinusoidal signals. (here noise is sinusoid and not white noise)
> > I'll give you a hint, look up adaptive line enhancer (ALE). It is a
> > LMS-based adaptive notch filter that makes its own reference. Make a
> > simulation of the ALE, then run it varying the update gain and see the
> > results as a function of the update gain. Then decide if LMS-based
> > filters work in this application.
>
> One approach to make the LMS work in this application is to add some
> artificial components into the signal. Those components could be used as
> a reference for system identification. The trick is how to make it
> efficient without introducing the significant artifacts.
>
> Vladimir Vassilevsky
> DSP and Mixed Signal Design Consultanthttp://www.abvolt.com

As long as the interference is a sinusoid, that's all that's needed.
The ALE will make it's own reference. Now,...... how well it works for
different types of signals (e.g., noise vs speech), I'm leaving up to
the OP to figure out when (s)he runs the simulations.
 |  Next  |  Last
Pages: 1 2
Prev: How to spec jitter for a digital NCO
Next: Literature/Link references?