Exponential sine sweep is a perfectly appropriate term to use, phase and frequency increase exponentially during the sweep.
It is configurable by the user on the Measure dialog, default is 300 ms. It only needs to be long enough for the response to decay to or below the noise floor.
A difficulty with trying to use an IR from a separate test is that the IR must align to much better than sample precision with the capture of the stimulus for which you wish to see the residual. Any misalignment would raise the residual as a result of a fractional sample time delay.
Convolution is done with FFT, doing it in the time domain is extremely inefficient unless the sequences are very short. Overlap techniques can be used to avoid very long FFTs.
If anyone wants the paper PM me.
Reading and learning. That link to the https://www.avnirvana.com/threads/fsaf-fast-subband-adaptive-filtering-measurement.13810/ is proving interesting.
Reading and learning. That link to the https://www.avnirvana.com/threads/fsaf-fast-subband-adaptive-filtering-measurement.13810/ is proving interesting.
I found similar problems a few years back trying to isolate low level signals from speakers. The technique was inspired by your hidden signal in the Audio DiffMaker software. A song (or other intentional audio signal) was hidden at a very low level in the stimulus signal, then played over a loudspeaker. Did that loudspeaker preserve or destroy the hidden low level signal? This was fairly easy to test with an electrical signal chain, but devilishly difficult once speakers, air, room and mics were in involved.
I am truly terrible at names. But a gentleman posted a FLAC file of what he measured. This is the spectral content of it. Looks to be power line harmonics as the main contributors. This is Musical Spectrum Analyzer plugin in Foobar2000. This is kind of what happens. Trying to extract all the noise that is generally masked.
The situation is much popular music is that distortion of the harmonics is the basis for the sound. This is even true in Classical music, adding beat frequencies and harmonics is a tried, tested and useful method to make sound more impressive. Reeds on pipe organs, over blown brass instruments.
Yes, practically all music instruments "pure tones" include overtones (harmonics) with sometimes higher amplitude than the fundamental note. This is why C from a flute sounds different than saxophone. Even individual instuments of same type have characteristic "tone" because relation of overtones varies. And eg. piano is easy to make sound a little different by opening the lid! And Stradivari et. al. legendary violins can be identified or emulated by studying their spectrum...
https://www.puremix.com/blog/musical-instruments
https://www.puremix.com/blog/musical-instruments
Last edited:
I see this now. I thought of micro moving the "model IR" with Mike's infamous Delta Spread Function but the computational load makes this expensive. Every speaker/mike measurement would need different twiddling.
You can still use naive deconvolution (using fancy FFTs of course) to get the 'model IR' from the original stimulus and then to get the residual.
There's something here just on the edge of my comprehension. We deconvolve and then truncate to get our "model IR". Then convolve "model IR" with stimulus for 'perfect' signal and subtract to get residual.
That means the residual must be sorta "in the bit we truncate". Angelo's method uses this 'truncated' bit to get the Harmonics.
Could we just convolve the truncated bit with our stimulus to get our residual? No problems with alignment then as we've already done the subtraction 😊
Wish I'd lern 2 reed, rite en kunt
Duu.uh! I was hoping for new 21st century supa dupa, no latency, zillion point methods. What size blocks do you use for overlap-and-save?
Last edited:
This is more difficult and certainly less fun than it would seem.
First there's the problem of choosing which program material to listen to in the test. There is literally an infinite number of choices, do you choose something busy, something with bits of silence between peaks, something ambient?. Something dynamic, voices, pipe organ, what?
Then there is the difficulty of doing the actual test on the ones you decide to try -- averaging some reasonably large number of takes (for both IR determination, and *of the track to be analyzed). You'd think you could set this up and go away till it's done a bunch of averages, but the acoustical world (unless you are in an environmentally quiet region of rural nowhere!) makes that impractical -- you need to stand by and monitor what's happening -- so you can start all over again! -- if a car drives by, a train in the distance comes through, a jet flies over, a neighbor starts mowing his lawn, the HVAC or refrigerator kicks on, etc., etc. Definitely do NOT pick one of your favorite tracks for the program material, you will get very sick of hearing it so many times, as you will of the signal used to collect the IR. And while you do have to stand by, you can't be in the same room because your body is a reflecting surface which will change effect if you move at all around there.
You'll find you won't be inspired to do too many of these tests acoustically! Doing them entirely electrically is much more practical to do, but of course checking loudspeakers is usually what you're after.
*This assumes that the DAC and ADC you are using are sample-locked to each other with common clocks, of course. Otherwise averaging takes of the program track isn't very practical without lots of messy and error-prone signal alignment processing.
Last edited:
Yeah I haven't actually tried this yet, it's just been very cool what mikets been posting. Wouldn't car passing by sound like a car in the residual? I get it it's tough to get quiet and repeatable, but I thought that this method was reasonably good with the stuff. Anyway, will try one day and see what you mean 🙂
Actual use case I'm gonna do is try whether series inductor makes audible difference with the woofer I have, since it's bit hard to hear difference directly I'm hoping to get the difference from residual and hopefully learn to detect barkhausen/hysteresis stuff just by ear. Effect of inductor or resistor is seen in normal distortion plots as well, so it's mainly just for self education and thus fun.
Actual use case I'm gonna do is try whether series inductor makes audible difference with the woofer I have, since it's bit hard to hear difference directly I'm hoping to get the difference from residual and hopefully learn to detect barkhausen/hysteresis stuff just by ear. Effect of inductor or resistor is seen in normal distortion plots as well, so it's mainly just for self education and thus fun.
Step 4 Analyse, Display and Listen to Residual This is the 3rd and final Process
Loadsa fancy pics. I use Farina sweep stimulus in my examples cos the results are predictable.
Spectrograms p38 fsaf.4.pdf Not sure Mike understands what is happening here.
Angelo's method effectively sweeps several narrow band filters; at the fundamental and at each hamonic. The 'model IR' is the output of the fundamental filter. When this is subtracted from the received signal, it removes a band around the fundamental. This is the deep blue trench in the residual spectrogram. Shows the ADAM F5 powered speaker is relatively Time Invariant (TI) for our 15s sweep (nice deep blue trench all the way up)
p11 Loudspeakers for AES.pdf
The first spectrogram is the residual from Angelo's method. The bright line below the blue fundamental trench at HF shows Focal PS130F is starting to compress towards the end of the 15sec sweep. By the time, the 'narrow band filter' reaches 1kHz & above, its voice coil has heated up so the fundamental isn't quite what the 'model IR'predicts. The subtraction isn't as total and the fundamental creeps into the residual giving the bright line.
The second spectrogram is FSAF ReLS (?) processing with the same Farina sweep. The 'model IR' is better matched at HF (less obvious HFbright line) but perhaps not quite as good at LF (less obvious deep blue trench)
Harmonics in both seem the same.
REW Residual #109,#110
DcibeL measures 2 speakers, first with Angelo's method, then uses the same Farina sweep with FSAF processing to get the REW Residual.
He shows REW Residual with FSAF is the 'same' as the Harmonic Distortions provided each harmonic is displayed at it's own frequency (REW's“Plot harmonics at the harmonic frequency” option) rather than at the frequency of the stimulus (as in traditional HD measurement)
#123 TNT corrects my misunderstanding of “Plot harmonics at the harmonic frequency” and provides me with probably the most important link between REW Residual and traditional HD.
#111 with Pink noise is just inaccurate. Noise methods become inaccurate with device distortion. The inaccuracies aren't so bad for response. #1& #34 But the subtraction for REW Residual highlights this inaccuracy.
Don't think any residual from a noise stimulus shows anything but this inaccuracy.
Loadsa fancy pics. I use Farina sweep stimulus in my examples cos the results are predictable.
Spectrograms p38 fsaf.4.pdf Not sure Mike understands what is happening here.
Angelo's method effectively sweeps several narrow band filters; at the fundamental and at each hamonic. The 'model IR' is the output of the fundamental filter. When this is subtracted from the received signal, it removes a band around the fundamental. This is the deep blue trench in the residual spectrogram. Shows the ADAM F5 powered speaker is relatively Time Invariant (TI) for our 15s sweep (nice deep blue trench all the way up)
p11 Loudspeakers for AES.pdf
The first spectrogram is the residual from Angelo's method. The bright line below the blue fundamental trench at HF shows Focal PS130F is starting to compress towards the end of the 15sec sweep. By the time, the 'narrow band filter' reaches 1kHz & above, its voice coil has heated up so the fundamental isn't quite what the 'model IR'predicts. The subtraction isn't as total and the fundamental creeps into the residual giving the bright line.
The second spectrogram is FSAF ReLS (?) processing with the same Farina sweep. The 'model IR' is better matched at HF (less obvious HFbright line) but perhaps not quite as good at LF (less obvious deep blue trench)
Harmonics in both seem the same.
REW Residual #109,#110
DcibeL measures 2 speakers, first with Angelo's method, then uses the same Farina sweep with FSAF processing to get the REW Residual.
He shows REW Residual with FSAF is the 'same' as the Harmonic Distortions provided each harmonic is displayed at it's own frequency (REW's“Plot harmonics at the harmonic frequency” option) rather than at the frequency of the stimulus (as in traditional HD measurement)
#123 TNT corrects my misunderstanding of “Plot harmonics at the harmonic frequency” and provides me with probably the most important link between REW Residual and traditional HD.
#111 with Pink noise is just inaccurate. Noise methods become inaccurate with device distortion. The inaccuracies aren't so bad for response. #1& #34 But the subtraction for REW Residual highlights this inaccuracy.
Don't think any residual from a noise stimulus shows anything but this inaccuracy.
LISTENING to REW Residual with inaccuracies … non-TI ness, EVIL noise stimuli bla bla
What loadsa people think is SOTA for speaker assessment.
Worth listening to? ABSOLUTELY!
It's the closest to being able to listen to non-Linear distortions with music.
Just bear in mind it's not always what you think it is, inaccurate, just noise, compression bla bla #158
More importantly, non-Linear speaker distortions are very far below Linear Distortions, frequency response, CDS (waterfalls), directivity, Room Interface Profile blabla, in the ranking of AUDIBLE speaker yuckies. see eg Intermodulation Distortion Listening Tests and the pontificating of the false prophets Olive & O'Toole
REW Residual doesn't highlight these, the MOST IMPORTANT speaker distortions.
The important Linear distortions don't even need subtraction to be audible to inexperienced listeners in DBLTs. These are the ones which make a speaker sound artificial.
One important Linear distortion that the REW Residual does pick up is compression (ie non TI ness) But I'm not sure listening to the residual demonstrates its true yuckiness on music, compared to listening to the complete system, warts & all.
FSAF a method to characterise (model, measure, bla bla) a system, sorta in 'real time' specially useful for Echo Cancellation and other fancy stuff. FSAF is probably SOTA for these applications.
Measuring speakers: As a modeling method, it should dream up 'model IR's with less noise than eg Angelo's method, in theory. In practice, this rarely happens. #152
In >260 pages of fsaf.1-4.pdf and “Loudspeakers for AEC”, Mike shows only one case where this is so, p49 fsaf.4.pdf, and maybe half a case on p39. In both these cases, Angelo's method produces a more than acceptable'model IR' some 8x faster.
For REW Residual, FSAF's ONLY role is to get an accurate 'model IR' which it needs cos of the subtraction.
But there are much simpler methods to get an accurate 'model IR' eg Angelo's method, and possibly the REW Residual too.
These might allow accurate 'model IR's to be applied to different stimuli for more accurate REW Residuals.
What loadsa people think is SOTA for speaker assessment.
Worth listening to? ABSOLUTELY!
It's the closest to being able to listen to non-Linear distortions with music.
Just bear in mind it's not always what you think it is, inaccurate, just noise, compression bla bla #158
More importantly, non-Linear speaker distortions are very far below Linear Distortions, frequency response, CDS (waterfalls), directivity, Room Interface Profile blabla, in the ranking of AUDIBLE speaker yuckies. see eg Intermodulation Distortion Listening Tests and the pontificating of the false prophets Olive & O'Toole
REW Residual doesn't highlight these, the MOST IMPORTANT speaker distortions.
The important Linear distortions don't even need subtraction to be audible to inexperienced listeners in DBLTs. These are the ones which make a speaker sound artificial.
One important Linear distortion that the REW Residual does pick up is compression (ie non TI ness) But I'm not sure listening to the residual demonstrates its true yuckiness on music, compared to listening to the complete system, warts & all.
FSAF a method to characterise (model, measure, bla bla) a system, sorta in 'real time' specially useful for Echo Cancellation and other fancy stuff. FSAF is probably SOTA for these applications.
Measuring speakers: As a modeling method, it should dream up 'model IR's with less noise than eg Angelo's method, in theory. In practice, this rarely happens. #152
In >260 pages of fsaf.1-4.pdf and “Loudspeakers for AEC”, Mike shows only one case where this is so, p49 fsaf.4.pdf, and maybe half a case on p39. In both these cases, Angelo's method produces a more than acceptable'model IR' some 8x faster.
For REW Residual, FSAF's ONLY role is to get an accurate 'model IR' which it needs cos of the subtraction.
But there are much simpler methods to get an accurate 'model IR' eg Angelo's method, and possibly the REW Residual too.
These might allow accurate 'model IR's to be applied to different stimuli for more accurate REW Residuals.
Last edited:
That's misrepresenting the process. When the response is deconvolved with the log sweep stimulus (or their FFTs are divided and invFFT'd) the sweep's group delay profile causes harmonics of the stimulus to be collected at increasingly negative times relative to the LTI response. The time regions corresponding to those harmonics can then be windowed and FFT'd to generate the harmonic responses, while the LTI region is windowed to generate the linear response.
Actually that's what effectively happens.
In da 90s, I was trying to dream up some way of measuring response & distortion in the shortest possible time for production testing. I started with various hardware swept filters then moved on to digital capture and an attempt to apply a zillion swept digital filters simultaneously on the captured record.
The "Aha!!" moment was when I realised, deconvolving with the stimulus was applying a Matched Filter to the Fundamental (for response) ... but also for each harmonic at different 'times'. You might argue that the filter shape isn't optimum but the bandwidth is so narrow that it doesn't really matter.
The British 'log' (or Italian 'exponential' 😊 ) sweep that allows this, was chosen fortuitously for constant S/N reasons.
At that time, the computing power, and particularly good A/Ds & D/As, were too $$$ for the many units I wanted for factory test so we developed ClioQC with Audiomatica. I introduced ClioQC to our China factories and it is now very much a Test Standard in China ... alas from Chinese copies
. Mauro Bigi of Audiomatica, probably still blames a certain Ozzie beach bum for this.When I emerged from the Ozzie bush in 2006, (?) I found even the cheapest computer had more than enough power to do this and good A/Ds & D/As were a dime a dozen.
Angelo visited said beach bum in Oz circa 2007 and was surprised to be shown Jurassic C code doing his stuff. I measured his supa $$$ Ambisonic microphone for him in my tin shed and dreamt up some digital filters on the spot to get it to sound good for the first time 😊
I'm quite happy to let Angelo take the credit as there was no way a beach bum could have persuaded AP, ARTA etc to make the method so readily available. Besides he gave me a lot of fancy gear, including a MOTU Traveler, new laptop, amps and 8 speakers for a Periphonic Ambisonic System which I've still got in boxes.
Today, I can make better measurements in my tin shed with Angelo's method and Jurassic DOS software ... than I could with Anechoic, B&K bla bla in da previous Millenium.
But Windows XP machines (that can run sensible DOS programmes are all at least 10 yrs old and nearly extinct
so I'll soon be forced to move into da 21st century and use REW 😊 with a new computer.
Last edited:
JohnPM, have you had a chance to ponder this?
A FSAF 'model IR' may not be long enough to do this (He may stop the modelling once it's long enough for a good 'model IR').
But you can certainly deconvolve a Farina sweep, music or other wide bandwidth signal (Don't use MLS or other noise) and get an IR as long as your stimulus.
The first bit is your LTI 'model IR'.
You convolve the rest (the truncated bit) with the stimulus to get your Residual.
Very efficient with zillion point FFTs. Only 2^20 = 1048576 points for 21.8sec 48k fs music
Last edited:
I would argue that the collection of harmonics is a welcome side effect of the particular group delay profile of a log sweep and not a filtering process.
It doesn't make sense to me. If the LTI response came from a log sweep, for example, harmonic distortions are shifted to negative time but intermodulation and noise have effects that are distributed throughout the IR and are not separable from the LTI component.
Member
Joined 2003
Don't mind me, user error is worked out. I've stopped posting anything here since it's just theory discussion.
I was hoping others would push the measure button and share results to help with some baseline comparison to aid in interpretation. But 9 pages in and I've yet to see a measurement result presented from users other than myself and @tktran303
Btw, ESS as a term has been in use for some 25+ years now. No one is going to start saying Angelo's method.
But don't you need multiple simultaneous frequency components to make IM? The longer the log sweep, the closer it gets to single frequency component driving the system at any time.
Some years back, I came up with a Farina-like test for some IM orders using two simultaneous sweeps (frequency separated by a fixed percentage, and sweeping backwards high to low). That brought out IM "impulse responses" in negative time (but also in positive time, so not as clean for frequency responses).
Noise can be improved by coherent averaging, compression by measuring at a lowest level. So I got my "clean" IRs by Angelo's sweep run at low-ish levels and averaged multiple times. Averaging gets limited by how TI you can trust the system to be over many long low level sweeps, though.
I have speakers that I am working on right now for a client. My only problem is my Quant Asylum rig is a bit of a pain to connect two microphones to. The Microphone preamplifier has one channel with powered phantom power. The second channel is basically for electret mics or to function as an instrumentation amplifier. I just received a USB powered phantom power supply. I'll rig up a cable between the two for signal.
The second stumbling block for me is figuring out how to set up REW correctly with this ASIO401 driver. I do like REW, so this is worth the time to do so. Then read up on how this new test method will work. Obviously we point mics and measure as normal. There's a few pages to read and digest.
I agree theory is important. But applied theory is where it get real no?
Member
Joined 2003
Perhaps you've over-complicated your own setup with that Quantasylum gear. A run of the mill USB audio interface is ideal for REW. ASIO driver and phantom power included 🙂
Last edited:
It's a solid setup. Allows a few things that are not so easy to do with a standard USB interface. It's new to me, and I am showing my age and slow integration of new toys 😉