Hi, Klaus; nice to hear from you, Julius. :)
As Julius points out, I also think that you'd need less demanding
thresholds: even for identical channels, the average using a one-pole lp
will oscillate and it might not be reliable if checking against the corner
case.
I'd go with something like this, but Julius also has a point about
phase-inverted mono signals. Also, you may want to normalise when you sum L
and R. It was necessary to guard against division by 0 too. I've included a
smoother to avoid clicks when switching.
I've tested that with your audio example and it's all mono except mono-left,
mono-right, and stereo. Is that correct?
import("stdfaust.lib");
avg(t, x) = fi.pole(p, (1 - p) * x) // 1-pole lowpass as average
with {
p = exp((((-2.0 * ma.PI) / t) / ma.SR));
};
var(t, x) = avg(t, (x - avg(t, x)) ^ 2); // variance
sd(t, x) = sqrt(var(t, x)); // standard deviation
cov(t, x1, x2) = avg(t, (x1 - avg(t, x1)) * (x2 - avg(t, x2))); //
covariance
corr(t, x1, x2) = cov(t, x1, x2) / max(ma.ma.EPSILON, (sd(t, x1) * sd(t,
x2))); // correlation
t = .5; // averaging period in seconds
correlate_meter(x,y) = x,y <: x , attach(y, (corr(t) :
hbargraph("corr",-1,1))) : _,_;
correlate_correct(t,l,r) = (l + r) * mSmoo + l * stSmoo , (l + r) * mSmoo +
r * stSmoo
with {
isMono = corr(t,l,r) > .999;
mSmoo = avg(.05, isMono);
stSmoo = 1.0 - mSmoo;
};
process = _,_ : correlate_meter : correlate_correct(t);
Ciao,
Dr Dario Sanfilippo
http://dariosanfilippo.com
On Mon, 9 Aug 2021 at 00:25, Julius Smith <julius.sm...@gmail.com> wrote:
> And of course I mean "cross-correlation coefficient"
>
> On Sun, Aug 8, 2021 at 3:22 PM Julius Smith <julius.sm...@gmail.com>
> wrote:
>
>> Hi Klaus,
>>
>> I am late to this (just read some of the thread with interest), and I
>> have a question: what do you mean by "1 > corr(t,l,r) > 0" ? It appears to
>> be "parsed" left to right, so that the 2nd ">" only sees "1>0" most (all?)
>> of the time, which is always true (1) of course (so no "else" activated).
>> Maybe you want something like "abs(corr(t,l,r)) > 0.95" ? (i.e., 95%
>> correlation deemed to be "panned mono"). I'm taking the absolute value
>> because I assume you don't care if the left channel is merely the negative
>> of the right (unless that's an accepted cheezy "stereoizer" of sorts).
>>
>> FYI, this is what we call a time-domain "normalized cross-correlation" or
>> "correlation coefficient" measurement (official buzzwords)
>>
>> Cheers,
>> Julius
>>
>>
>> On Sun, Aug 8, 2021 at 10:07 AM Klaus Scheuermann <kla...@posteo.de>
>> wrote:
>>
>>> Dear Dario,
>>>
>>> cool, your corr function gives me the desired results. At least when
>>> feeding it to a meter.
>>>
>>> Here is my test audio which contains vocals in mono-mid, mono-left,
>>> mono-right, mono-half-left, mono-half-right, stereo:
>>> https://cloud.4ohm.de/s/y9oZzqFGyrZT5ej
>>> For mono-mid, mono-half-left, mono-half-right it shows 1.
>>> For mono-left, mono-right it shows 0.
>>> For stereo it shows values between 0 and 1.
>>>
>>> I would like to detect mono signals that are not exactly in the middle
>>> and put them there. Stereo signals should be unchanged.
>>>
>>> My code is here, but for some reason it does not work correctly.
>>> Especially when corr shows 0, ba.if does not go to the else-path.
>>>
>>> import("stdfaust.lib");
>>> avg(t, x) = fi.pole(p, (1 - p) * x) // 1-pole lowpass as average
>>> with {
>>> p = exp((((-2.0 * ma.PI) / t) / ma.SR));
>>> };
>>> var(t, x) = avg(t, (x - avg(t, x)) ^ 2); // variance
>>> sd(t, x) = sqrt(var(t, x)); // standard deviation
>>> cov(t, x1, x2) = avg(t, (x1 - avg(t, x1)) * (x2 - avg(t, x2))); //
>>> covariance
>>> corr(t, x1, x2) = cov(t, x1, x2) / (sd(t, x1) * sd(t, x2)) : _ ; //
>>> correlation
>>> t = 0.5; // averaging period in seconds
>>> correlate_meter(x,y) = x,y <: x , attach(y, (corr(t) : hbargraph("corr",
>>> -1,1))) : _,_;
>>> correlate_correct(t,l,r) = ba.if(1 > corr(t,l,r) > 0 , l, (l+r)) , ba.if
>>> (1 > corr(t,l,r) > 0 , r, (l+r));
>>> process = _,_ : correlate_meter : correlate_correct(t);
>>>
>>> Am I doing the ba.if wrong?
>>>
>>> Thank s very much,
>>> Klaus
>>>
>>>
>>> On 04.08.21 18:25, Dario Sanfilippo wrote:
>>>
>>> I had implemented a few statistics function a while back, kindly taken
>>> from Wikipedia, and they seem to produce the expected values mentioned on
>>> the webpage. I hope that these can be useful.
>>>
>>> Ciao,
>>> Dr Dario Sanfilippo
>>> http://dariosanfilippo.com
>>>
>>> import("stdfaust.lib");
>>> avg(t, x) = fi.pole(p, (1 - p) * x) // 1-pole lowpass as average
>>> with {
>>> p = exp((((-2.0 * ma.PI) / t) / ma.SR));
>>> };
>>> var(t, x) = avg(t, (x - avg(t, x)) ^ 2); // variance
>>> sd(t, x) = sqrt(var(t, x)); // standard deviation
>>> cov(t, x1, x2) = avg(t, (x1 - avg(t, x1)) * (x2 - avg(t, x2))); //
>>> covariance
>>> corr(t, x1, x2) = cov(t, x1, x2) / (sd(t, x1) * sd(t, x2)); //
>>> correlation
>>> ph0 = os.phasor(2.0 * ma.PI, 200);
>>> red = sin(ph0) + .35 * sin(ph0 * 3.0) + .91 * sin(ph0 * 5.0);
>>> blue = sin(ph0) + .5 * sin(ph0 * 3.0) - .5 * sin(ph0 * 5.0);
>>> red1 = sin(ph0) + sin(ph0 * 3.0);
>>> blue1 = sin(ph0) - sin(ph0 * 3.0) / 3.0;
>>> t = 1.0; // averaging period in seconds
>>> process = (red , blue : corr(t)) , (red1 , blue1 : corr(t));
>>>
>>>
>>>
>>> On Wed, 4 Aug 2021 at 16:52, Klaus Scheuermann <kla...@posteo.de> wrote:
>>>
>>>> Thanks Giuseppe,
>>>>
>>>> I checked it out, but somehow it still does not give me the desired
>>>> result...
>>>> I did some more research and found this, which indicates that it can be
>>>> done with arctan more easily.
>>>>
>>>> The way this is done on phase (correlation) meters in audio equipment
>>>> is rather simple:
>>>>
>>>> Phase = arctan(L/R)
>>>>
>>>> With phase of 45 or 225 = 1, and phase of 135 and 315 (-45) is -1.
>>>>
>>>> Essentially, the Y Axis is the L, and the X axis is the R. The phase is
>>>> simply the polar angle of the vector between the two.
>>>>
>>>> This type of meters will show 1 if the signal is mono, and -1 if the
>>>> left and right are perfectly phase inverted.
>>>>
>>>> Notice however, that phase meters of this type also account for the
>>>> magnitude in the polar coordinates. So:
>>>>
>>>> Magnitude = (L^2 + R^2)^1/2
>>>>
>>>> Thus the actual meter display is a normalised version of:
>>>>
>>>> Correlation = Phase * Magnitude
>>>>
>>>> I'm not sure that satisfies your requirements, but this answers the
>>>> question in the subject.
>>>>
>>>> So I transfered this to faust, but it still behaves weired...
>>>> import("stdfaust.lib");
>>>> phase(l,r) = (l/r) : aa.arctan;
>>>> magnitude(l,r) = (l^2 + r^2)^1/2;
>>>> correlate(l,r) = phase(l,r) * magnitude(l,r);
>>>> correlate_meter(x,y) = x,y <: x , attach(y, (correlate : hbargraph(
>>>> "corr",-1,1))) : _,_;
>>>> process = _,_ : correlate_meter: _,_;
>>>>
>>>> Any ideas?
>>>>
>>>> Klaus
>>>>
>>>> On 03.08.21 14:48, Giuseppe Silvi wrote:
>>>>
>>>> Hi Klaus,
>>>> The filters are necessary to obtain a -1 +1 range, I think.
>>>>
>>>> import("stdfaust.lib");
>>>>
>>>> correlate(l,r) = l*l ,r*r , l*r : par(i,3, si.smooth(0.9)) : sqrt, sqrt, _
>>>> : *,_ : /;
>>>> correlate_meter(x,y) = x,y <: x , attach(y, (correlate :
>>>> hbargraph("corr”,-1,1)));
>>>>
>>>> process = correlate_meter;
>>>>
>>>> Try playing with the si.smooth coefficient.
>>>>
>>>> best,
>>>> giuseppe
>>>>
>>>>
>>>> On 3 Aug 2021, at 14:09, Klaus Scheuermann <kla...@posteo.de>
>>>> <kla...@posteo.de> wrote:
>>>>
>>>> Could it be something like this?
>>>>
>>>> (according to the 'correct' algorithm in
>>>> https://www.beis.de/Elektronik/Correlation/CorrelationCorrectAndWrong.html
>>>> )
>>>> import("stdfaust.lib");
>>>>
>>>> correlate(l,r) = l*l ,r*r , l*r : sqrt, sqrt, _ : *,_ : / :_;
>>>> correlate_meter(x,y) = x,y <: x , attach(y, (correlate :
>>>> hbargraph("corr",-1,1))) : _,_;
>>>>
>>>> process = _,_ : correlate_meter: _,_;
>>>>
>>>> I am not sure about the lowpass filters though. Maybe not needed in the
>>>> digital domain?
>>>>
>>>> Also, my code only returns -1 or 1 while it should be returning a range of
>>>> -1 and 1, right?
>>>>
>>>> The correlation is either expressed in % from -100% to +100% or as the
>>>> correlation factor, which ranges from -1 to +1. Note that due to the
>>>> correlation algorithm the level of both signals does not matter, i.e., it
>>>> does not influence the measured result.
>>>>
>>>> When a mono source is used for a stereo signal both stereo channels will
>>>> be +100% correlated. When e.g. in a stereo signal both channels contain
>>>> completely different signals, e.g. left (and only left) is the trumpet and
>>>> right (and only right) is the guitar these stereo channels will be 0%
>>>> correlated. With a third instrument appearing in both channels, the
>>>> correlation will be somewhere between 0 and +100%.
>>>>
>>>>
>>>> Ideas?
>>>> Danke :)
>>>> Klaus
>>>>
>>>>
>>>>
>>>>
>>>> On 03.08.21 12:48, Klaus Scheuermann wrote:
>>>>
>>>> Hello List,
>>>>
>>>> I just wondered, if anyone has implemented a stereo audio correlation
>>>> meter/analyser in faust?
>>>>
>>>> If yes - great!
>>>> If no - I have another project :)
>>>>
>>>> Here is what I found about the algorithm(s):
>>>> https://www.beis.de/Elektronik/Correlation/CorrelationCorrectAndWrong.html
>>>>
>>>>
>>>> I never learned analog electronic schematics, but it seems it should not
>>>> be extremely hard to transfer to faust.
>>>>
>>>> Cheers, Klaus
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Faudiostream-users mailing list
>>>> Faudiostream-users@lists.sourceforge.nethttps://lists.sourceforge.net/lists/listinfo/faudiostream-users
>>>>
>>>> _______________________________________________
>>>> Faudiostream-users mailing
>>>> listFaudiostream-users@lists.sourceforge.nethttps://lists.sourceforge.net/lists/listinfo/faudiostream-users
>>>>
>>>> _______________________________________________
>>>> Faudiostream-users mailing list
>>>> Faudiostream-users@lists.sourceforge.net
>>>> https://lists.sourceforge.net/lists/listinfo/faudiostream-users
>>>>
>>> _______________________________________________
>>> Faudiostream-users mailing list
>>> Faudiostream-users@lists.sourceforge.net
>>> https://lists.sourceforge.net/lists/listinfo/faudiostream-users
>>>
>>
>>
>> --
>> "Anybody who knows all about nothing knows everything" -- Leonard Susskind
>>
>
>
> --
> "Anybody who knows all about nothing knows everything" -- Leonard Susskind
>
_______________________________________________
Faudiostream-users mailing list
Faudiostream-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/faudiostream-users
