Actually, if done properly, Angelo's method can be as fast as you want.
A fast sweep just means your frequency resolution (smoothing) and noise are worse. It's a good way to demonstrate how non Linear Time Invariant speakers are. Us speaker guys just find it convenient to pretend speakers are LTI.
And BTW, if the dynamic mike has the same frequency response as the condensor, its signal will arrive at the same time.
If you want to be pedantic, because the dynamic will usually have less LF than the condensor, you might say its signal will arrive first ... but that ignores stuff like da Hilbert Transform and Minimum Phase which will apply to all 'measurement' or 'simple' mikes.
Last edited:
With Angelo's method, the sweep rate actually sets the smoothing. eg a 1sec sweep actually sweeps a 1/1sec = 1Hz bandwidth (sync) filter so 1 Hz smoothing. ie longer Farina sweeps have LESS smoothing.
What REW appears to do is apply 1/24 8ve smoothing to this. It would explain the noise floor rising with frequency. BTW, this is the correct way to display noise on a LOG frequency scale like all these here. If a LIN frequency scale is used for the display, then a constant bandwidth smoothing is appropriate for noise.
Yes. Angelo's method is at least 3dB better (usually a LOT better) at this party trick. I'm not sure if your TDS method can measure THD too? MLS is about equivalent to TDS for response & noise immunity but can't do THD.
D
Deleted member 375592
The "tap bandwidth" is set by FFT/IFFT number of bins. You can augment 1 sec with as much silence as you want to get whatever frequency resolution. But Lars is correct - longer sweeps have more energy transmitted, therefore better SNR, ~sqrt(T), higher dynamic range, etc. Noise floor risings are in part due to Farina's outdated regularization. You can do better, but his general approach is fine with me.
I don't think you can assume the MK255+MV210 behaves like the B&K mike and amp. You certainly can't assume this behaviour for da electret 'measurement' mikes.
For those planning to make their own Measurement mikes, take heed of Bill Waslo's caveats about Electrets & their built in FETs. Most of them use '3 terminal' capsule and operate the FET as a Source Follower. For these, Bill's caveats are exact.
But there is another 'high spl' mode to operate such a capsule that appears to give perhaps 3dB less THD and maybe 3dB more spl before overload.
This is to use it in a Charge Amp configuration. I describe this in SimpleP48.pdf I think this is suitable as a 'Measurement Mike' for response but not for THD for which I'd use a slightly more elaborate circuit. Also in MicBuilders are Eric Benjamin's measurements of what Bill describes and also Simon Stefaneli's measurements of SimpleP48RCA. You have to join
Last edited:
Adding silence may show more bins but doesn't actually increase frequency resolution. It's the length of real data that counts.
Agreed
No. Angelo's method will show flat spectrum for White Noise. I have an example in Zephyr.doc in MicBuilders Files. That's also where Eric Benjamin's report on Panasonic WM-61_Rev5.pdf is.
mikets42, I'm struggling to grok our FSAF method by going through your Mathwork page but my single brain cell is lacking. Have you published this in a peer reviewed journal where hopefully they will have asked you to use words of 1 syllable? 😊
Is AEC Adaptive Echo Cancellation?
Last edited:
OK, I've gone through this thread a Third time at great cost to my single remaining brain cell 😱
tktrans303, Am I right in saying NONE of the cheapo measurement mikes you have tested are anywhere near your 1/2" B&K THD measurements?
BTW, I'm really suspicious of ALL the measurements you show in #1. Are they all done with REW?
I think you said your 1/2" B&K measurements were done with REW ... hence the 1/24 8ve smoothing on the Farina sweeps.
tktrans303, Am I right in saying NONE of the cheapo measurement mikes you have tested are anywhere near your 1/2" B&K THD measurements?
BTW, I'm really suspicious of ALL the measurements you show in #1. Are they all done with REW?
I think you said your 1/2" B&K measurements were done with REW ... hence the 1/24 8ve smoothing on the Farina sweeps.
D
Deleted member 375592
No, I did not. I can not make it any shorter than 200 pages, and no journal would accept that long a text 🙂 BTW, what else would you expect from a guy who started speaking at 15? However, I did re-write the introduction for R3.0. Do you want to be the first reader?
Yes, it all started with AEC, when I discovered that I could do fine and fast adaptive filtering. However, I needed B&W 800 class speakers to produce low enough LTI distortions, which made it all meaningless. Then, I started to dig into loudspeaker distortions and ended up on this forum.
I concur with you: no sense wasting any time on cheap electret mics.
Frequency resolution is not defined by the length of excitation. You can send a 1ns-long delta function and observe a RIR 10 sec long. It all goes back to the radars in the early 50s. They could spread a pulse with SAW filters which were limited by the length of the xstal (20mm then), with a Rayleigh wave speed of ~1500 m/s. The interpulse period could be much longer.
Last edited by a moderator:
Agreed.
What puzzled me is how the Klippel NFS setup measures distortion. Initially I thought that it was related to the default microphone that Klippel sales engineers recommend for their NFS. I thought Erin might benefit from the
Mic301E IEPE 1/4" set.
But in the most recent measurement of another PTT6.5X mid-woofer in a similar sized cabinet, it is clearly able to resolve H2 -60dB and H3 around -70dB, so it's not that:
This is class leading harmonic distortion performance in the bass down to 35Hz.
But what is the cause of those spikes ~155Hz and 578Hz, I could only guess.
Last edited:
Farina method (exponential sine sweep):
1. the Signal to Noise Ratio of the raw signal is independent of the sweep length: longer sweep means more signal energy but also more noise energy by the same ratio (assuming stationary noise)
2. same goes for the deconvolved impulse response (IR)
3. When we window the IR we chose the time-frequency trade off: a short time window cuts noise but leaves most of the IR energy at the expense of the spectrum being smoothed.
4. for a given window length then a longer sweep improves the SNR : we keep a smaller fraction of the noise energy.
4. the exp sweep signal has a pink spectrum . assuming white noise,the measured response has a noise density rising 3dB/octave.
5. by shrinking the window inversely proportional to the analysed frequency we flatten the noise again and get a constant relative resolution bandwidth (eg 1/48 octave). this is done in a recent update to REW.
6. and that’s all i have to say about shrimp fishing
1. the Signal to Noise Ratio of the raw signal is independent of the sweep length: longer sweep means more signal energy but also more noise energy by the same ratio (assuming stationary noise)
2. same goes for the deconvolved impulse response (IR)
3. When we window the IR we chose the time-frequency trade off: a short time window cuts noise but leaves most of the IR energy at the expense of the spectrum being smoothed.
4. for a given window length then a longer sweep improves the SNR : we keep a smaller fraction of the noise energy.
4. the exp sweep signal has a pink spectrum . assuming white noise,the measured response has a noise density rising 3dB/octave.
5. by shrinking the window inversely proportional to the analysed frequency we flatten the noise again and get a constant relative resolution bandwidth (eg 1/48 octave). this is done in a recent update to REW.
6. and that’s all i have to say about shrimp fishing
D
Deleted member 375592
@Irisbo ESS / MLS / and other "perfect sequence" methods avoid Fisher matrix inversion by using delta-correlated basis as excitation, so they are computationally fast, especially with FFT processing. For any p=2 method as LS/BLUE, SNR grows with sqrt of excitation time for uncorrelated stationary observation noise.
@tktran303 this is far from class-leading performance. Moreover, the graphs lack the confidence intervals for each harmonic, without them you can't say what is and what is not. Moreover, a driver can not physically produce such graphs - they are artifacts of the measurements, Moreover, representing distortions per harmonics is based on the implicit assumption of smoothness, so the Fourier expansion is quite close to eigenvector (KL) expansion; when that is not the case (and it is not) PCA has nothing to do with harmonics.
@tktran303 this is far from class-leading performance. Moreover, the graphs lack the confidence intervals for each harmonic, without them you can't say what is and what is not. Moreover, a driver can not physically produce such graphs - they are artifacts of the measurements, Moreover, representing distortions per harmonics is based on the implicit assumption of smoothness, so the Fourier expansion is quite close to eigenvector (KL) expansion; when that is not the case (and it is not) PCA has nothing to do with harmonics.
A cardioid pattern is nothing more than just two monopoles added together with change in delay (phase) and level.
This can be done with microphones as well.
Either directly or in post.
In fact, this has been done and used for decades and described as a basic configuration in any engineering book about microphones.
(So I am confused about the question here?)
Even off the shelf solutions exists like the Zoom H2N etc
(Probably not suitable for freq resp measurements)
The proximity effect with cardioid mics is a direct effect of how these cardioid mics physically and mechanically work.
Omni mics don't have these issues.
that is not true. you might think of TDS where the smoothing is a function of the sweep rate?
You may like to make a cardioid out of two omnis.
The B&K Sound Intensity wands are just such a beast but it's easy enough with a couple of electret omnis. You'll find the resultant device shows proximity too. IIRC, the B&K literature / manuals show this clearly but why don't you try this for yourself.
There's a small bit of signal processing you need to do for dis cardioid but it can be easily done analogue
Mike, could you please explain some of these fancy terms for us wid only 1 brain cell? eg ESS, 'perfect sequence', fisher matrix inversion, p=2, LS/BLUE, KL expansion, PCA etc.
If some of yus other DSP gurus can chip in with words of 1 syllable, this would be appreciated too 🙂
You are conflating two different 'windows'.
There's the 'measurement' window which is directly related to the sweep length. A longer sweep and hence longer 'measurement' window gives you better S/N. You can check this easily by measuring with different sweep lengths and looking at the tail of the IR. Not sure if REW can do this without adding extra processing.
Windowing the IR sets your frequency resolution. But the S/N of this 'display' window has the same S/N as above ie independent of the size of your 'display' window. So here you are right.
But this 'display' window isn't the same as the 'measurement' window. The 'measurement' window IS related to sweep length and gains S/N with longer sweeps.
no conflation : window is applied to the IR. not to be confused with the the length of the stimulus or observation time (which limits the how long IR you can estimate). it’s the ratio between raw IR and windowed IR that buys you a gain in SNR
D
Deleted member 375592
ESS=exponential sine sweep
perfect sequence = usually tri-state, the extension of MLS, most used in CDMA, a family of sequences having delta-function auto- and cross-correlation properties. see https://www.sciencedirect.com/science/article/abs/pii/S1434841116313632 for further reading.
The Fisher aka Information matrix is sum(x(t)*x(t)'/sigma(t)^2), the base of Least Squares, x(t)'*h=y(t); t=1:N; X={x(1)',x(2)',...,x(N)}; Y={y(1),y(2),...,y(N)}; Fisher matrix=X'*X; X'*X=X'*Y; Y=inv(X'*X)*X'Y; When you use perfect sequences, X'*X becomes a unity matrix, and for EES - almost, so you don't have to invert it, you need only calculate X'*Y which, when expanded, becomes simply a correlation of x and y.
p=2: you can choose yourself which p-norm you minimize, {sum(err^p)}^(1/p), p->inf = Chebyshev, p->1 = median, p=2 -> Least Squares (LS) which is also Best Linear Unbiased Estimator (BLUE).
KL expansion, aka eigenvector expansion, close to Principal Component Analysis (PCA), is usually referred to as Karhunen–Loève transform, which is the basis of functions that have the fastest decay rate for the (here) distortions, i.e., you will need least number of "KL harmonics" to describe distortions with best precision. If the sequence of harmonics does not converge quickly, your basis (Fourier for THD) is not appropriate, you need another basis that converges faster. For example, figure 4 in Lars' https://purifi-audio.com/blog/tech-notes-1/this-thing-we-have-about-hysteresis-distortion-3 describes crossover distortions (actually identical to Class B amplifier's distortions) contains 2 spikes, which are quite close to delta-functions. THD calculated on this signal will have a neverending sequence of identical odd harmonics. What is then THD, if it grows with the number of harmonics accounted for? You can state any number to describe it - if you omit the notion of bandwidth. The proper KL transform shall use only a few coefficients to describe such crossover distortions regardless of the bandwidth of measurement equipment.
perfect sequence = usually tri-state, the extension of MLS, most used in CDMA, a family of sequences having delta-function auto- and cross-correlation properties. see https://www.sciencedirect.com/science/article/abs/pii/S1434841116313632 for further reading.
The Fisher aka Information matrix is sum(x(t)*x(t)'/sigma(t)^2), the base of Least Squares, x(t)'*h=y(t); t=1:N; X={x(1)',x(2)',...,x(N)}; Y={y(1),y(2),...,y(N)}; Fisher matrix=X'*X; X'*X=X'*Y; Y=inv(X'*X)*X'Y; When you use perfect sequences, X'*X becomes a unity matrix, and for EES - almost, so you don't have to invert it, you need only calculate X'*Y which, when expanded, becomes simply a correlation of x and y.
p=2: you can choose yourself which p-norm you minimize, {sum(err^p)}^(1/p), p->inf = Chebyshev, p->1 = median, p=2 -> Least Squares (LS) which is also Best Linear Unbiased Estimator (BLUE).
KL expansion, aka eigenvector expansion, close to Principal Component Analysis (PCA), is usually referred to as Karhunen–Loève transform, which is the basis of functions that have the fastest decay rate for the (here) distortions, i.e., you will need least number of "KL harmonics" to describe distortions with best precision. If the sequence of harmonics does not converge quickly, your basis (Fourier for THD) is not appropriate, you need another basis that converges faster. For example, figure 4 in Lars' https://purifi-audio.com/blog/tech-notes-1/this-thing-we-have-about-hysteresis-distortion-3 describes crossover distortions (actually identical to Class B amplifier's distortions) contains 2 spikes, which are quite close to delta-functions. THD calculated on this signal will have a neverending sequence of identical odd harmonics. What is then THD, if it grows with the number of harmonics accounted for? You can state any number to describe it - if you omit the notion of bandwidth. The proper KL transform shall use only a few coefficients to describe such crossover distortions regardless of the bandwidth of measurement equipment.
We don't need live audio on the fly.
This can be done 100% in post, so I would go the easy route.
Also to keep the original signal.
In fact it can be done with just one mic.