Thank you - that makes more sense. When I measure distortion I did notice that the software needed to switch to a sweep, rather than MLS - which I did not think much of, since I got the same response - as you also point out.
If, using music as a test-signal, resemble a kind of multi-tone distortion test, then it could be cool, because suddenly we see the potential capacity of a given speaker system, since a speaker drive isn't perfectly linear - like a test sweep. So we could hope to see how different speakers could maybe handle more or less complex music, with the least amount of distortion.
Interesting to see if this brings out a way to maybe clarify how little we can even detect.... like how much is enough for our hearing?
That should be
"Lipshitz & Vanderkooy who helped develop MLS in Jurassic times, would think us very silly to try to measure HD with noise"
Member
Joined 2003
Not so. Thanks DcibeL for answering Dmitrij_S and providing references.
This used to happen with old sweep methods eg B&K 2010 and 2307.
It does not happen with Angelo's method. You can vary the sweep rate from minutes to less than a second and still get the same response altered only by the theoretical 'smoothing' and noise due to the short 'measurement'.
Loadsa old B&K articles on the subject which will give you insights into why Angelo's method is the theoretically fastest method to measure response or distortion for a given accuracy & noise floor.
That's why we should call Angelo's method by its proper name. Not all exponential (or log 😊 ) sine sweep methods are equal.
Thanks Kravchenko. Please measure a speaker.
I think measuring a good interface or amp gets very close to the inherent inaccuracies of floating point calculation. Before yus DSP gurus jump on me, it's only an opinion based on investigating the self dithering properties of IIRs which is why I prefer them to EVIL FIRs.
And please use the same Farina sweep at the same level for your FSAF measurement.
No. If you use a longer Farina sweep, its noise immunity increases. A Farina sweep even 10x slower than your stepped sine will still have better than 10x noise immunity.
Not sure how REW does Angelo's method but this noise immunity is only limited by the convenient size of your FFT to do 'long' measurements.
When we started doing this in da 70s, we were surprised to find that only the BBC had attempted this before. Today, only da false prophets Olive & O'Toole have done similar stuff. But they reference us so must be OK 😊
Simulation and Investigation of Doppler Distortion
Audibility of Amplifier Clipping
Intermodulation Distortion Listening Tests
The last paper is the important one as it has nearly all the results of our listening tests on various distortions in the appendix. It's the one that's most quoted by da false prophets Olive & O'Toole 😊
I think this makes you a DSP guru by diyAudio standards 😊
I've read fsaf.1.pdf in detail cos Mike sent me a draft of a re-write asking for comments. Alas, by the time I'd dreamt some up and got back to him, he'd left the forum.
I'm compiling a list of TLAs which he uses and hope to get explanations suitable for my single remaining brain cell.
I note he says But it’s NOT as “fast” as FFT where all opportunities have been exploited in his Preface on p2 😊 which of course is what Angelo's method uses.
Only skimmed through fsaf.2 - 4 and also Loudspeakers for AEC.
If we assume his implementation of FSAF is good (confirmed by the fact that it gives good response similar to several other methods both more & less efficient) the big question is how he obtains his residual and whether it shows the (non) Linear part of LTI. This is not clear from any of the 5 PDFs. "Loudspeakers for AEC" and fsaf.4 seem to present it as a fiat accompli. I probably need to overload my one remaining brain cell studying his MATLAB code.
When Kravchenko, Tranh and DcibeL show that FSAF Total Distortion replicates old fogey type measurements, I'll be forced to this 😲
I think Bill Waslo shares my reservations about what the FSAF residual actually represents.
Yes please, Kumar.
If you know of these papers, or any others which explain subband filtering accurately and simply in the public domain, preferably simple enough for my single brain cell, please post links. My naive skimming suggests FSAF is good for 'modelling' and Mike even uses some tricks that I like for dreaming up digital filters 😊 but gotta unnerstan da TLAs to confirm this.
I no longer have MATLAB and the Signal Processing Toolbox and, more importantly, the Help files in the Toolbox. I'm only a senile pre10 DSP guru from da Jurassic 80s & 90s.
Last edited:
Member
Joined 2003
No, FSAF in general, it doesn't measure HD specifically, the chart doesn't discern between 2nd harmonic, 3rd harmonic, etc. It is "total distortion" of all types including harmonics. In order to separate harmonics, the source frequency of the distortion must be known, so a sine wave reference as a single frequency "fundamental" reference must be used.
Here's an interesting comparison. A "good speaker" was tested, with fairly low level of 80dB/1m. Same sine sweep was run through normal ESS ...err. "Angelo method".. as well as through FSAF.
Here, the harmonics are shown in an unconventional "show harmonics at harmonic frequency" because I think it makes this comparison a lot clearer. It also highlights the uselessness of a THD vs frequency chart, since THD for a given frequency doesn't indicate whatsoever which harmonics are being energized, it's a completely useless graph without individual harmonics as well, so might as well turn that plot off entirely as it doesn't contain any information that isn't available without it.
So key here is to forget about THD white line, and focus on just the peak (largest) harmonic for any given frequency. If you follow that comparison, I actually think the FSAF provides a rather comparable result. Main difference as well is that FSAF includes a higher noise floor as it is listening to all frequencies at once regardless of the sine sweep stimulus. This test is completed in my office, a computer runs in the background so it's not the quietest space in the world.
I increased SPL by 6dB, 86dB/1m
Here's an interesting comparison. A "good speaker" was tested, with fairly low level of 80dB/1m. Same sine sweep was run through normal ESS ...err. "Angelo method".. as well as through FSAF.
Here, the harmonics are shown in an unconventional "show harmonics at harmonic frequency" because I think it makes this comparison a lot clearer. It also highlights the uselessness of a THD vs frequency chart, since THD for a given frequency doesn't indicate whatsoever which harmonics are being energized, it's a completely useless graph without individual harmonics as well, so might as well turn that plot off entirely as it doesn't contain any information that isn't available without it.
So key here is to forget about THD white line, and focus on just the peak (largest) harmonic for any given frequency. If you follow that comparison, I actually think the FSAF provides a rather comparable result. Main difference as well is that FSAF includes a higher noise floor as it is listening to all frequencies at once regardless of the sine sweep stimulus. This test is completed in my office, a computer runs in the background so it's not the quietest space in the world.
I increased SPL by 6dB, 86dB/1m
Member
Joined 2003
Here's another good speaker. Same comparison as above, sine sweep used for both HD and FSAF, at same level.
Member
Joined 2003
For another comparison, here's the same speaker as above, but instead of a sine sweep, I tested with 80dB/1m pink noise. Quite a different result indeed!
Reasonably close correlation between Sweep method in REW, and the FSAF method in REW, with the Measurement Sweep signal as the stimulus.
I'm not surprised that Pink (or White) Noise appears higher. In my view, it's a tough test because it's all the frequencies playing, all at once.
Tougher than a multi-tone test using 12 equally spaced tones between 20Hz and 20KHz, or even 12 tones per octave
eg.
Reference:
https://www.klippel.de/products/rd-system/modules/mton-multi-tone-measurement.html
Please attach some audio recordings of the FSAF Residual when available.
I'm not surprised that Pink (or White) Noise appears higher. In my view, it's a tough test because it's all the frequencies playing, all at once.
Tougher than a multi-tone test using 12 equally spaced tones between 20Hz and 20KHz, or even 12 tones per octave
eg.
Reference:
https://www.klippel.de/products/rd-system/modules/mton-multi-tone-measurement.html
Please attach some audio recordings of the FSAF Residual when available.
Last edited:
Member
Joined 2003
Noise signal means all frequencies all at once, and distortion includes intermodulation of all frequencies all at once, and peaks in signal that are 10dB greater than sine sweep of the same level. I have to think there's a good reason that some sweep is not a signal option in REW for FSAF, it's the wrong stimulus for the intent of this measurement. The only intermodulation that occurs in a sine sweep is intermod of the harmonics which are already very small signal.
This is what is so interesting with FSAF - to get everything "in between" the harmonics. Your following comparisons are very interesting. It shows an increased "noise" at louder levels even if one could assume that the background noise if fairly constant. This I assume is the "hash" and "hish" one can hear as a tail of modulated noise at times - I think this is one of the worst type of distortion and I would very much like to see it gone. And now we can measure it. To do also good ol THD+N meauremns in pairs with FSAF will be goof practise probably...
Many thanks for your presented measurements.
//
Greater signal-to-noise ratio of the sweep sine doesn't mean the better noise immunity. It is well known, that the MLS technique randomize the phase spectrum of the output signal which is not correlated with the MLS inputs sequence. In other words, the phase spectrum of any external noise/disturbance signal (white, impulsive noise or any other type) will be randomized. This allows you to reduce the external noise effect on the impulse response by applying averaging of successive MLS bursts.
You can do this with Angelo's method too. I do this cos my Jurassic programmes can't do zillion point FFTs. The limit is set by the reverb of your room (cos residuals of earlier sweeps) but you can concatenate 10 sweeps in a domestic room and more in an anechoic.
Last edited:
Thanks for this DcibeL.
Any REW gurus here that can explain this?
Alas, this makes it very difficult to see if THD(H2..9) is correct. I'm suspicious of this in REW. If I read you correctly, I need to shift the Red 2nd harm curve up an 8ve and the Orange 3rd up 3x in freq. etc. to get to the traditional HD display.. But then, THD(H2..9) would appear to be LESS than some of the harmonics when it should be more. (It should sum Power).
Any REW gurus here that can explain this?
Are you saying that the FSAF residual is akin to your "harmonics at harmonic frequency" display? Your results certainly suggest that.
But how is this (single??) "all frequencies at once" noise figure displayed with varying level against frequency??
DcibeL, Tranh, Kumar and any other DSP and MATLAB gurus here, I'm trying to figure out what the FSAF residual display is. I know I should be able to do this by inspection of the MATLAB code but it might take me the rest of the Millenium to grok something sensible.
Could you gurus please comment on my probably naive, understanding (guess?) of what Michael is doing.
Could you gurus please comment on my probably naive, understanding (guess?) of what Michael is doing.
- Use FSAF with arbitrary stimulus to produce a long IR which models the supposedly LTI transfer function of your DUT (including room responses bla bla) This produces an 'accurate' response which we know corresponds to other methods for IR & response. Loadsa buzzwords and TLAs which I don't unnerstan at present but we know it gets 'good' response.
- Convolve this IR with the stimulus to get what should be the perfect result if the transfer function is LTI (ie has no distortion, compression bla bla)
- Subtract the expected perfect result from the actual result to get the residual.
- Analyse the residual to get the display.
- But what are the analysis parameters to get this display?
REW has a setting that can show distortion either which frequency it is trigger by or where the unwanted contribution sticks out their ugly heads 😉
Click the Controls wheel.... Plot the harmonics at the harmonic frequency... (I never get what the harmonic frequency is off the two 🙂 )
//