>The xcorr method to obtain TDoA is all good if the signals do not contain
>too much noise. If the signals are corrupted by noise you better use one
>of the Generalized Cross Correlation (GCC) methods like PHAT, SCOT, HT,
>etc.
>
>Fortunately there is a pretty good, and easy to use, implementations of
>GCC in the Matlab file exchange page:
>
>
>http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=8581&objectType=file
>
>
>Here is a little example of how I used it. I am not expert so don't
expect
>much of this code. Replace xnoisy and ynoisy with your data obtained
from
>the sensors and set Nfft and Fs to your needs. You can read the help
files
>for pwelch and cpsd for details. For more information on GCC methods
simply
>google for "Generalized Cross Correlation".
>
>
>%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>% Estimate the spectra.
>%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>[Sxx, Fxx] = pwelch(xnoisy,[],[],Nfft, Fs);
>[Syy, Fyy] = pwelch(ynoisy,[],[],Nfft, Fs);
>[Sxy, Fxy] = cpsd(xnoisy,ynoisy,[],[], Nfft, Fs);
>
>%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>% Obtain the Generalized Correlation
>%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>[corr1, t1] = GCC('unfiltered',Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>[corr2, t2] = GCC('scot', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>[corr3, t3] = GCC('roth', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>[corr4, t4] = GCC('phat', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>[corr5, t5] = GCC('cps-m', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>[corr6, t6] = GCC('ht', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
>
>figure(1);
>subplot(321);
>plot(t1,corr1);
>%axis([-1 1 -1 1]);
>[m,k] = max(corr1);
>k = t1(k);
>title('Unfiltered GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));
>
>subplot(322);
>plot(t2,corr2);
>%axis([-30e-6 30e-6 -1.5 1.5]);
>[m,k] = max(corr2);
>k = t2(k);
>title('SCOT GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));
>
>subplot(323);
>plot(t3,corr3);
>%axis([-1 1 -1 1]);
>[m,k] = max(corr3);
>k = t3(k);
>title('ROTH GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));
>
>subplot(324);
>plot(t4,corr4);
>%axis([-1 1 -1 1]);
>[m,k] = max(corr4);
>k = t4(k);
>title('PHAT GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));
>
>subplot(325);
>plot(t5,corr5);
>%axis([-1 1 -1 1]);
>[m,k] = max(corr5);
>k = t5(k);
>title('cps-m GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));
>
>subplot(326);
>plot(t6,corr6);
>%axis([-1 1 -1 1]);
>[m,k] = max(corr6);
>k = t6(k);
>title('HT GCC');
>text(1e-4,0.5,strcat(num2str(k),'sec'));


I USED YOUR CODE BUT I AM HAVING A PROBLEM.
MATLAB SAYS THAT THERE IS SOME KIND OF ERROR AND THIS IS WHAT IT SAYS:

??? Error using ==> unknown
Matrix dimensions must agree.

Error in ==> GCC at 166
R = Pxy .* W;

Error in ==> TDOA at 16
[corr1, t1] = GCC('unfiltered',Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);

DOES ANYBODY KNOW HOW TO SOLVE THE PROBLEM? 

>> It looks like more samples are desired in the delay domain 

Exactly that: Zero padding in the frequency domain gives me oversampling
in the time domain, so I can resolve the delay between the signals with
higher resolution.
Sorry if that was unclear, thanks for clarification.

Cheers

Markus

On Aug 9, 3:37 pm, Jerry Avins <j...@ieee.org> wrote:
> mnentwig wrote:
>
>    ...
>
> > You'll need to oversample, otherwise the accuracy is too bad, given the
> > velocity of sound (one sample at 44.1ksps corresponds to 7 mm)
> > Easily done inside the FFT by zero-padding - split FFT(sig1) in the middle
> > and insert for example zeros with 3 x the length for oversampling of four.
>
> Do I understand correctly that you propose to add information by zero
> padding?
>
>    ...
>
> Jerry
> ...

It looks like more samples are desired in the delay domain to resolve
a peak position to a finer scale than uninterpolated peak-picking can
provide. Zero-padding FFT interpolation does that. In Matlab, adding
the zeros in the middle  of the signal can correctly implement that.
Why would anyone call that adding information?

Dale B. Dalrymple
http://dbdimages.com

"Jerry Avins" <jya@ieee.org> wrote in message 
news:QYidnayPv_apCibbnZ2dnUVZ_tjinZ2d@rcn.net...
> mnentwig wrote:
>
>   ...
>
>> You'll need to oversample, otherwise the accuracy is too 
>> bad, given the
>> velocity of sound (one sample at 44.1ksps corresponds to 
>> 7 mm)
>> Easily done inside the FFT by zero-padding - split 
>> FFT(sig1) in the middle
>> and insert for example zeros with 3 x the length for 
>> oversampling of four.
>
> Do I understand correctly that you propose to add 
> information by zero padding?

Well, at any rate, it will be very clean information, 
utterly devoid of anything that might be mistaken for signal 
;-)

mnentwig wrote:

   ...

> You'll need to oversample, otherwise the accuracy is too bad, given the
> velocity of sound (one sample at 44.1ksps corresponds to 7 mm)
> Easily done inside the FFT by zero-padding - split FFT(sig1) in the middle
> and insert for example zeros with 3 x the length for oversampling of four.

Do I understand correctly that you propose to add information by zero 
padding?

   ...

Jerry
-- 
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯

Hi,

some quick shooting-from-the-hip concept: 
I assume for simplicity that there is a single scratch sound, and the
recorded time window is much longer (silence before and after, means I can
use cyclic FFT without most of the worries)


- Calculate IFFT(FFT(sig1) .* conj(FFT(sig2)))
Explanation in "numerical recipes in C", FFT section (cyclic
convolution).
Search for the peak in the (complex) result. Its location is related to
the time difference between sig1 and sig2. 

You'll need to oversample, otherwise the accuracy is too bad, given the
velocity of sound (one sample at 44.1ksps corresponds to 7 mm)
Easily done inside the FFT by zero-padding - split FFT(sig1) in the middle
and insert for example zeros with 3 x the length for oversampling of four.


If noise is too high (the peak isn't clear enough), use a filter that
matches the expected amplitude spectrum of your sound (after the FFTs on
sig1 and sig2, multiply with H(f)).

BTW, are two mics sufficient? Without absolute time reference, wouldn't I
need a third mic? Anyway, it doesn't matter, the same code that gives me
the time difference between mic 1 and 2 will also work for mic 1 and 3.


The FFT processing approach is overkill on a cosmic scale for a product
implementation, but it should do for a concept demo.

Hope that's useful.

-Markus

Thanks for all the replies...

What I have rigged up so far, is an MDF surface (representing the
'mouse/touch' pad) and using two condensor microphones placed on two
corners of the pad to record a finger moving across the MDF surface which
essentially is passive audio location.

I know humidity etc will affect how it will sound. Seeing as I'm leading
towards trying to do a 'proof of concept' type research, I've gotten rid
of humidity issues by using talc on the surface, so (if you can imagine) a
slight hissing noise as the finger slides across it.

In terms of how loud it is, you can distinctively hear it on the recorded
wave files. There is a significant amount of noise in the signal however,
mainly at 50Hz and some other random frequencies due to my computer being
rather loud. I have taken my board to different locations and computer
labs (less noisy environments) and recorded a few samples there as well.

I have found the frequency band of interest and filtered out all other
frequencies as best I can and have done this (simply put):

[cov, lags] = xcov (wav_L, wav_R, 'coeff');
delays = lags / fs;
plot (delays, cov); 

With recordings of localised sound (ie finger sliding around a single spot
on the MDF board), I took the peak values in the plot and recorded them.
There is a noticable trend with a graph of the delays are plotted (ie from
Left most point to Right most point) but the amount of error that I was
getting made me think twice about using the results as the delay could end
up positive instead of being negative in certain cases. I feel this might
be my recording method and I'm worried that this might make it even more
complicated to implement. 

Even after filtering most of the noise out, assuming white noise is
equally present in both channels, and considering my computer CPU is
sitting to the right of the pad when recording, could it affect my delay
plots enough to cause this error? (I'm pretty sure my CPU induces noise at
50 and 22 kHz only (pwm fan controller))

The amount of challenges involved in this project has gotten me over the
top and i'm afraid that i might not be able to complete it at all...



Thanks ryujin_ssdt, I will give that code a try..

The xcorr method to obtain TDoA is all good if the signals do not contain
too much noise. If the signals are corrupted by noise you better use one
of the Generalized Cross Correlation (GCC) methods like PHAT, SCOT, HT,
etc.

Fortunately there is a pretty good, and easy to use, implementations of
GCC in the Matlab file exchange page:


http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=8581&objectType=file


Here is a little example of how I used it. I am not expert so don't expect
much of this code. Replace xnoisy and ynoisy with your data obtained from
the sensors and set Nfft and Fs to your needs. You can read the help files
for pwelch and cpsd for details. For more information on GCC methods simply
google for "Generalized Cross Correlation".


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Estimate the spectra.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
[Sxx, Fxx] = pwelch(xnoisy,[],[],Nfft, Fs);
[Syy, Fyy] = pwelch(ynoisy,[],[],Nfft, Fs);
[Sxy, Fxy] = cpsd(xnoisy,ynoisy,[],[], Nfft, Fs);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Obtain the Generalized Correlation
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
[corr1, t1] = GCC('unfiltered',Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
[corr2, t2] = GCC('scot', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
[corr3, t3] = GCC('roth', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
[corr4, t4] = GCC('phat', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
[corr5, t5] = GCC('cps-m', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);
[corr6, t6] = GCC('ht', Sxx, Syy, Sxy, Fs, length(xnoisy), Nfft);

figure(1);
subplot(321);
plot(t1,corr1);
%axis([-1 1 -1 1]);
[m,k] = max(corr1);
k = t1(k);
title('Unfiltered GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

subplot(322);
plot(t2,corr2);
%axis([-30e-6 30e-6 -1.5 1.5]);
[m,k] = max(corr2);
k = t2(k);
title('SCOT GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

subplot(323);
plot(t3,corr3);
%axis([-1 1 -1 1]);
[m,k] = max(corr3);
k = t3(k);
title('ROTH GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

subplot(324);
plot(t4,corr4);
%axis([-1 1 -1 1]);
[m,k] = max(corr4);
k = t4(k);
title('PHAT GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

subplot(325);
plot(t5,corr5);
%axis([-1 1 -1 1]);
[m,k] = max(corr5);
k = t5(k);
title('cps-m GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

subplot(326);
plot(t6,corr6);
%axis([-1 1 -1 1]);
[m,k] = max(corr6);
k = t6(k);
title('HT GCC');
text(1e-4,0.5,strcat(num2str(k),'sec'));

cheers,
Horacio

>On Aug 6, 6:27 pm, "nickch" <nic...@slingshot.co.nz> wrote:
>> >Will the environment be very clean, as in the only sound you hear
>> >is only the mouse click?
>>
>> No, I assume that the environment will be pretty noisy and have taken
>> steps to identify the frequency of interest (sliding finger) and
filter
>> out the rest.
>>
>> >Did you try using xcov or xcorr in Matlab yet?
>>
>> I have not tried xcorr yet, but I am using xcov to get the covariances
and
>> lag values and plotting that to find the peak which should give me the
>> delay in the signal (assuming what my supervisor says is correct).
>>
>> >I'm having a hard time figuring out what you have tried, what you
>> >don't
>> >know how to try, and what you do not know how to try.
>>
>> What I would like to know is: with a stereo wave file loaded into
matlab
>> (wavread), how would I accurately determine and possibly plot the time
>> difference of arrival between the two channels?
>>
>> Thanks
>
>Have you actually tried using xcov/xcorr to see if you can possibly
>get a good estimate?  What did you find?  Have you tried it in the
>presence of additional noise sources?
>
>See, I still can't tell what you have tried, and what you have found
>from your trials and experiments.
>
>At this point I tend to agree with Rune, I doubt that this approach
>will work in anything but a very synthetic environment.  Have you
>at least tried looking at what kind of noise you pick up at the
>sensors?
>
>Can you separate the sound of a mouse click from {everything else
>that sensors pick up}?
>
>Do you understand how finding the peak in the xcov/xcorr gives
>you the time delay of arrival?
>
>

On Aug 6, 6:27 pm, "nickch" <nic...@slingshot.co.nz> wrote:
> >Will the environment be very clean, as in the only sound you hear
> >is only the mouse click?
>
> No, I assume that the environment will be pretty noisy and have taken
> steps to identify the frequency of interest (sliding finger) and filter
> out the rest.
>
> >Did you try using xcov or xcorr in Matlab yet?
>
> I have not tried xcorr yet, but I am using xcov to get the covariances and
> lag values and plotting that to find the peak which should give me the
> delay in the signal (assuming what my supervisor says is correct).
>
> >I'm having a hard time figuring out what you have tried, what you
> >don't
> >know how to try, and what you do not know how to try.
>
> What I would like to know is: with a stereo wave file loaded into matlab
> (wavread), how would I accurately determine and possibly plot the time
> difference of arrival between the two channels?
>
> Thanks

Have you actually tried using xcov/xcorr to see if you can possibly
get a good estimate?  What did you find?  Have you tried it in the
presence of additional noise sources?

See, I still can't tell what you have tried, and what you have found
from your trials and experiments.

At this point I tend to agree with Rune, I doubt that this approach
will work in anything but a very synthetic environment.  Have you
at least tried looking at what kind of noise you pick up at the
sensors?

Can you separate the sound of a mouse click from {everything else
that sensors pick up}?

Do you understand how finding the peak in the xcov/xcorr gives
you the time delay of arrival?

On 4 Aug, 14:58, "nickch" <nic...@slingshot.co.nz> wrote:
> Hi,
> I've been thrown into a project that involves determining the location of
> a sound source and tracking its movements. I have very little matlab
> experience but have been trying to cope. This site has been very useful.

Your courage is impressive. As I write this, my radio is playing
by my PC, playing loud so I can hear it over the noise from the
neighbour's kids who are playing soocer just outside the open
window by my desk. If you want to track the sounds of a mouse
(BTW, I use a mouse mat with mine) in such an environment,
you certainly have your work cut out for you.

This reminds me of a project proposal I once saw, for a
seismic sonar for land mine detection in agricultural
areas. The person who hatched the idea had never worked
with farming or gardening, as he did not know that soil
contains stones, lumps of clay, roots and so on which
would destroy any attempts to single out land mine
targets.

> The general idea for the project is implementing a finger mouse where you
> have an array of microphones that detects the sound emitted by a finger
> sliding across a surface and emulate that as a mouse on a computer.

A *finger* sliding across a surface? How do you model the
sound source representation of greasy soft tissue sliding
across a hard surface? How loud is it? What are its
characteristics? How robust is this source with respect
to skin humidity? Air humidity? Surface roughness?
Surface hardness?

Rune