Swept sine deconvolution
Hello Katja,
that is great news!!!
I was looking into the scaling factor today but get stopped by work stuff and I am planning to have a look tonight.
I will also test the patch as well.
If you want I can do a quick real world test here at work as I have all the equipment.
In my first attachment there was a sort of implementation of reproduce and recording of the signal (if I remember well), would this be of help??
Bassik
Also forgot to mention that looking around for another reason I found this software made in Java:
http://www.hometheatershack.com/roomeq/
It does IR measurements as well.
Cheers
Bassik
Included some recording-, save- and load-routines to the log sweep deconvolution, and done a first real world test in my room. Dreadful standing waves oh boy.
Status quo patch attached.
Some weighted amplitude normalisation should be done on the sweep response. Now I have done an easy peak normalisation, but this will always lead to IR's having too low power.
Anyway, it is already possible to see the effect of the log sweep method (by the way this works with a linear sweep as well): non-linear distortion and pollution is shifted left of time zero and thus separated from the impulse response proper. Use vertical zoom to see this.
It would be useful to include an elementary wave editor for trimming. I have now done this with Audacity. Hardoff has done a wave editor in Pd, we could use a subset of it and include vertical zoom.
Katja
http://www.pdpatchrepo.info/hurleur/logsweepdeconv-test.pd.zip
@bassik said:
I found this software made in Java:
It's great and it does all you need plus more.
Our Pd patch may not grow to such a professional tool, if only because it would require years of developement. But from an educational perspective a Pd implementation may be more interesting. I remember very well how IR measurement was demonstrated by our dsp teacher in Sonology class. We went through the mathematics, and I was excited about how you can expand and compress time dimension with help of a sweep.
With our Pd patch, the whole process can be visualised. At any point in it, you can test how (im)precise your method and result is. You can then decide whether it is good enough for your purpose, or should be improved. You can read Ben Saylor's [partconv~] code to see how fast convolution is done in practice. The benefits of open source. I see good reason to continue our log sweep deconv project.
Katja
Hello Katjav,
The new patch is great, it works well and already does more than I was expected to reach in such a small amount of time.
I would like to do some tests and comparison with a commercial software with the same functionality but I have just realised that the laptop with the software installed is not available until monday; nevermind.
I have two questions on the patch:
there is a strange behaviour in PD, if i turn on dsp the patch doesn't allow me to move the horizontal slides or better it doesn't visualize the movements while with dsp off there is no problem. It is a bit of a problem as you don't see the computation progress slider moving.
Do you know why?
No error messages and the last PD installed.
would you please expand a bit more about weighted amplitude normalisation?
in the case of IRs normalising the main peak of the file will make all the IR information readable, or probably I am missing the point here.
so regarding the to-do list:
Multichannel: it is quite important it could be multichannel input (i.e. ambisonic mic or stereo pairs) while it is much less important as multichannel output (1 signal would be common to all speakers you are testing, no need for different signal).
fft analysis of IR: great idea but would be great to derive also additional acoustic parameters, I will provide maths for all the parameters we need to implement (ISO 3382 part 1 and part2, sorry I need to send the papers privately as I would get in trouble if I share on the web).
Also I would like to add in the future, a-format to b-format encoding (http://pcfarina.eng.unipr.it/Public/B-format/A2B-conversion/A2B.htm).
This is why 4 capsules tetrahedral mics can be used to analyse 3D soundscapes including directional analysis (along 3 axes) and understand issues in room such as room modes directions, flutter echoes and first reflections direction among others.
I am glad you have done a clean-up of the scaling constant calculation as in the previous patch it was a bit confusing to me.
I will have a more in depth play around tomorrow or friday...i am also in the middle of moving house...:(
Thank you
Bassik
@bassik said:
would you please expand a bit more about weighted amplitude normalisation? in the case of IRs normalising the main peak of the file will make all the IR information readable
The sweep itself has RMS 1/sqrt(2) as it is purely sinusoidal. Deconvolved, it translates as a unit pulse which, as a filter, would have neutral amplitude effect. A sweep response however, fluctuates with peaks and dips over the frequency range. When peak-normalised, it has only dips and it's RMS will be lower than 1/sqrt(2), assuming it still has sinusoidal character. This also translates to an IR with low power. When used as a filter, such an IR can reduce overall signal output substantially. In my first tests, by some 10 dB or more. See what I mean?
Therefore I think there should be an option to somehow normalise RMS of the sweep response / impulse response. I have no clue how this is handled in regular analysis software. Maybe in frequency domain, after the IR is trimmed to it's useful length?
Katja
@bassik said:
there is a strange behaviour in PD, if i turn on dsp the patch doesn't allow me to move the horizontal slides
This quirk has popped up a few times recently on Pd forum
http://puredata.hurleur.com/sujet-4893-number-slider-boxes-freeze
The deconvolution patch is in no way optimized. The [Scope~] showing sweeps is running at highest possible refresh rate. The deconvolution subpatch does crazy lots of fft's continuously. These things need switches to turn them off occasionally.
Katja
Hello Katja,
thank you very much for your answer.
Unfortunately in the next couple of weeks I would not be able to be present on the web due to work commitments and moving house with my wife (no internet at home...).
So I will try to check the forum and have a look at the amplitude normalisation but I am not sure I will have enough time to study it properly.
Sorry about that.
I will get in contact as soon as everything will be sorted.
CHeers
Bassik
Hello Katja,
Fortunately I had a conversation with a colleague today about IR normalisation for another reason and probably we got our answer.
In simulation softwares (Acoustics 3D models simulations), the auralisation module when deriving an IR and then the convolution of the anechoic file, applies a gain at the two stages and stores the info in the file.
In the case of the IR it applies a gain in order to have the maximum peak digital value (0.9999999999) and then consequently the same gain to all the rest of the IR (24-bit integer).
When doing the auralisation it applies another normalization gain to prevent clipping of the auralised file and takes into account that auralisation is made at 16-bit integer.
a simple program for the IR normalisation (converting the IR to PCM) can be:
Variable definitions
FP(N) = floating point impulse response
PCM(N) = linear impulse response
M = sample length
Algorithm
MAX = 0
For N = 1 to M
If abs (FP(N))> Max then MAX = abs(FP(N))
Next N
GAIN = (2^16)/MAX
For N = 1 to M
PCM(N) = FP(N) * GAIN
Next N
Store GAIN
As you can see the gain is arbitrary and depends on the IR (or the recorded sweep response) and there isn't a general rule as far as I know.
IN our case we should do these normalisation of the file and store somewhere (metadata in the wav file) the gain applied that would be useful in future applicaitons.
IN the case of multichannel IRs this can be a bit more complex but one thing at time is better.
Hope this helps
Bassik
Hello Katja,
one other comment:
gain formula should be:
GAIN = (2^15)-1/MAX for a 16-bit file.
or GAIN = (2^23)-1/MAX for a 24-bit file
CHeers
Bassik
@bassik said:
In the case of the IR it applies a gain in order to have the maximum peak digital value (0.9999999999) and then consequently the same gain to all the rest of the IR (24-bit integer).
Ah, so normalisation of the IR is done in such a way that the resolution of the integer format is maximally employed. That makes sense.
And then for convolution of the anechoic signal, a second normalisation factor should be applied, in such a way to prevent clipping. This is the difficult part. How to compute this factor from an arbitrary IR? I don't see how it could be computed from the IR array itself, since the convolution output magnitude depends on input frequency. But I can imagine a test method: convolve a sweep with the IR and find the maximum output, then scale down to have maximum output 0.999999 when expressed in floats. Is that how it is done?
All taken together, lots of routines to be written in Pd. Oh yeah, there is something else: for proper analysis, the impulse response of the test set (speaker&mic) should be factored out. Taking close-mic sweep response, invert magnitude spectrum, compensate, etc. Is this regularly done in IR measurement practice? If so, how?
Katja
@katjav said:
How to compute this factor from an arbitrary IR? I don't see how it could be computed from the IR array itself, since the convolution output magnitude depends on input frequency.
Oops, this is of course very simple to find out in frequency domain since the trimmed impulse response will easily fit in one FFT frame. This thread is now getting full with my stupid remarks, sorry.
Katja
I have now written an object class [chirp~], generating linear or logarithmic chirps calculated with 64 bit float precision. It has user parameters for start / stop frequency and chirp length. It has two outlets, one for the chirp and one for the inverse chirp.
The chirps from this object have neglegible numerical error, as opposed to the earlier chirp-generating patches. It is a pity that you can no longer see how the chirps are calculated, without reading the C-code. But this is the only road to quality IR measurement within Pd.
Attached is a zip containing chirp~.c, chirp~.pd_darwin, and chirp~-help.pd. If you happen to be on OSX, you can check [chirp~] rightaway. Otherwise you may compile the class for your platform, or wait till I have done that on someone else's computer.
Katja
Hello Katjav,
I am finally back even if still with some problems as I don't have internet at home and work is manic....
Anyway let's put everything in the right order:
[chirp~]: I am not on mac but I can test it on mac at home but it would be good to compile it for all platforms...do you have any reference where I can understand how to do the compiling?
also when you consider linear sweep, are you considering a different inverse filter which only the time reversal of the test signal?
In attachment there is an extract of the CATT acoustics manual on calibration of IR and convolution; It is related to audio convolution for auralisation simulations but the same principle can be applied to our purposes to calibrate IR and convoluted responses.
Katjav wrote: Oh yeah, there is something else: for proper analysis, the impulse response of the test set (speaker&mic) should be factored out. Taking close-mic sweep response, invert magnitude spectrum, compensate, etc. Is this regularly done in IR measurement practice? If so, how?
Normally in IR measurement practices with Swept sine, the convoluted result will appear in the software with all the non-linearities on its left hand side. In all common software (as far as I know) the removal of the non linearities is done by hand just by changing the start point of the result IR and by windowing the end part of it....after this you start analyzing all the parameters you want to derive from the IR.
I hope I answered your question.
also please have a look at this paper: http://pcfarina.eng.unipr.it/Public/Papers/238-NordicSound2007.pdf
from page 13 to 24 it explains all the issues that can occurs with sweep sine method.
I will try to start the tests soon and to try to find more info in literature.
Hope this helps
Thank you
Bassik
http://www.pdpatchrepo.info/hurleur/CATT_acoustics_IR_calibration.pdf
Thanks Bassik for pointing to the Farina article. Particularly interesting is the section on Kirkeby inverse filters for compensating spectral deficiencies in measurement equipment.
Meanwhile, I have made a patch to inspect characteristics of [chirp~] output in detail, with horizontal and vertical zoom options on the IR plot, and spectrum magnitude and phase plots. Real-signal chirps can never have a completely white spectrum. It turns out that logarithmic chirps are quite sensitive to parameter settings, while linear chirps are less so. Therefore, a log chirp may have considerable ripple near start and stop frequency, but I have also found settings which induce low ripple (+/- 0.3 or 0.4 db). It may have been coincidence that I stumbled upon such favourable settings. There is probably no simple rule to derive them numerically.
Because the log chirp is so sensitive to parameter settings, it needs special attention. For an IR measurement tool, we could provide 'approved' chirps in stereo wave files (32 bit floating point) together with their inverse. Chirp generator / tester and IR measurement tool would then be two different Pd root patches which can be developed independantly.
So, the chirp generator / tester is attached to this post. Again, there is only an OSX binary for [chirp~] (same one as before). Sorry. Building on other platforms is easy pie if you have developer tools installed and know how to operate them, but otherwise, don't spoil your time on it. I'll do it later on a friend's computer.
Katja
Hello Katjav,
hope you are well.
I have performed some test on a real case and I have to say that despite the first 2 measurements worked very well, after that the deconvolution process stopped to work properly.
Attached there is one of the wrong results I have obtained from the soft.
you can clearly hear that instead of the IR you get the inverse filter at the same time where the IR should be.
Is it a user problem?
Is it a multi measurement issue with the software?
I will investigate further before any conclusion.
thanks
Bassik, there may be some confusion. This last patch chirp~test.pd is only a patch to test the qualities of the generated chirp itself.
I am kind of swallowed by details of the method. Instead of quickly producing a working IR measurement tool, I am now trying to refine the method, starting with refining the chirp itself. Last time I mentioned the sensitivity for parameter settings. Now I have found a systematic method to optimize a chirp. The formula is different from the regular ones as found in texts. This has lead to yet another chirp object [expochirp~] (yes it's exponential, not logarithmic, also see Farina's 2007 article).
The regular exponential chirp formula's make a chirp start at sine phase to avoid a step, but they do not care about the phase at the end of the function, therefore it may well end with a step and produce excessive ripple at the high spectrum end. This is traditionally relieved by windowing, or by manually cutting the chirp at the last zero-crossing as Farina suggests. My new formula computes the chirp in such a way that it will start and stop in sine phase. No need for windowing or cutting, and the ripples are minimized.
Still there will always remain ripples in a chirp's spectrum, and I am now trying to produce inversion-filters to clean up a system's measured IR. For the high spectrum end, this is fairly well possible. But in the low spectrum end, where it is more important, the ripple character of exponential chirps is such that conventional methods fail. I did find some solution but it's not without compromise. Therefore I am investigating alternative methods.
I am sorry that proceedings are slow now. But my hope is to contribute to a good quality tool in the end. I'll post details on the new chirp formula soon.
Katja
Hello Katjav,
I haven't used the latest chirp~test patch but the previous version of the soft with teh old chirp computation.
So it was supposed to work at least in terms of recording the chirp and deconvolve it; but as i said it needs more testing as it was just a one off mesurements set the one i performed.
I am sorry as well I haven't contributed much in teh last time but I am full of stuff at work.
Anyway I will post a comparison between the PD measurement and a professional soft measurement in teh same positions soon.
Also I am not able to use your latest chirp~test patch as I am on windows.
thank you for keeping the study alive!!!
HEllo Katjav,
As promised I have performed some comaprisons between the IR measured by EASERA software and the one done with the PD patch we are developing.
As said, I have used the old chirp computation patch and have obtained useful IRs only for the first 2 measurements.
comparison in RT results are in the attached file showing:
Interesting is to see that EASERA is not able to compute the dB SPL of the wav file as it doesn't have any reference of the microphone calibration.
I need to investigate more if I can insert this info for the wav file so for the moment the waterfall plots are not comparable (EASERA measurements are in dB SPL while PD measurements in dB FS).
Anyway this is just a start and you can see that the RT measurements are quite similar apart for the 125 Hz results.
Hope this helps
More to come
Cheers
Hey Bassik, thanks so much for your test results.
Apart from the 125 Hz the found RT are comparable indeed. But the magnitude plots of the second case? Should M0002_S01_R02.etm be comparable to Test 2b proc.wav? Even without a reference, the figure should show a similar shape because it is in deciBels, no?
But well, at least it shows that the Pd patch is not complete nonsense. We're on the right track.
I wonder how the magnitude of a full IR is computed in EASERA. For a 20 second chirp the IR it is at least 40 seconds, anticausal and causal part taken together. Or is the analysis limited to the causal part? No, I would think that full IR means the whole thing with the left side included. Pd can not do such long FFT's, 131072 points is the maximum. Probably the magnitude could be computed as an average of several overlapping frames. But the phase information is lost then.
I consider writing an FFT object for Pd that operates on a named buffer, and works with 64 bit floats internally. Maybe it is then possible to do larger FFT's. This is also important for calculating inverse filters (correction filters), since the full IR spectrum is sometimes needed for that. There is also something wrong with [fft~] and [rfft~] phase interpretation: it does not do a centered FFT. So with an IR centered you get these alternating phases instead of a zero-phase representation. Therefore, complex multiplication of spectra will give wrong results. Quite annoying. A new FFT object should have an option to select centered and non-centered FFT so you could work well with centered IR, and with causal part only. Hmm, plans....
Katja
Oops! Looks like something went wrong!
