Sum,
It is difficult to know where to start, as some of your statements demonstrate a profound lack of knowledge regarding measurement, whatever people may have told you about the depth of your understanding. I'll simply reiterate the basics, as that may be useful for others reading the thread, and provide some common language we can fall back on when you "get done ripping this impulse testings down to the almost nothing they are".
Measurement equipment, in this context, tries to determine the transfer function of the system it tests, which is a description of how the system modifies signals passed through it. That transfer function can be described in the frequency domain through the system's magnitude and phase responses, or in the time domain through its impulse response. They are two sides of the same coin, and the FFT lets us move back and forth from one to the other.
To determine the transfer function the measurement equipment sends a reference signal through the system and captures the system's output. In modern acoustic test equipment that signal is often a logarithmically swept sine wave, which has some useful properties which I won't go into in this post. The system's transfer function can be worked out by carrying out an FFT on the reference signal, to determine its magnitude and phase response, then doing the same on the system's output. Dividing the FFT of the output by the FFT of the input yields the transfer function of the system, how it affects the magnitude and phase of the spectrum of any signal passed through it. An inverse FFT of that transfer function yields the system's impulse response, which shows how it affects any time signal passed through it.
That, in essence, is all there is to measuring a transfer function. Different reference signals have different properties that make them more or less suited to particular measurement environments, but the process can be applied to noise signals, sweeps, music or almost anything else, although the quality of the results will depend very much on the characteristics of the signals used. For those interested in a deeper understanding of measuring transfer functions, covering nearly all the techniques that are or have been used in acoustics, this paper fits the bill: http://www.anselmgoertz.de/Page10383/Monkey_Forest_dt/Manual_dt/aes-swp-english.PDF.
John
It is difficult to know where to start, as some of your statements demonstrate a profound lack of knowledge regarding measurement, whatever people may have told you about the depth of your understanding. I'll simply reiterate the basics, as that may be useful for others reading the thread, and provide some common language we can fall back on when you "get done ripping this impulse testings down to the almost nothing they are".
Measurement equipment, in this context, tries to determine the transfer function of the system it tests, which is a description of how the system modifies signals passed through it. That transfer function can be described in the frequency domain through the system's magnitude and phase responses, or in the time domain through its impulse response. They are two sides of the same coin, and the FFT lets us move back and forth from one to the other.
To determine the transfer function the measurement equipment sends a reference signal through the system and captures the system's output. In modern acoustic test equipment that signal is often a logarithmically swept sine wave, which has some useful properties which I won't go into in this post. The system's transfer function can be worked out by carrying out an FFT on the reference signal, to determine its magnitude and phase response, then doing the same on the system's output. Dividing the FFT of the output by the FFT of the input yields the transfer function of the system, how it affects the magnitude and phase of the spectrum of any signal passed through it. An inverse FFT of that transfer function yields the system's impulse response, which shows how it affects any time signal passed through it.
That, in essence, is all there is to measuring a transfer function. Different reference signals have different properties that make them more or less suited to particular measurement environments, but the process can be applied to noise signals, sweeps, music or almost anything else, although the quality of the results will depend very much on the characteristics of the signals used. For those interested in a deeper understanding of measuring transfer functions, covering nearly all the techniques that are or have been used in acoustics, this paper fits the bill: http://www.anselmgoertz.de/Page10383/Monkey_Forest_dt/Manual_dt/aes-swp-english.PDF.
John
Skimming through the IASYS manual, it appears to work similar to ARTA in FR2 (2-channel frequency response) mode. This is just a loopback of my sound card but it shows a periodic noise signal (similar to pink noise but not random) and a plot of FR, phase and coherence without going through the impulse intermediate step. Delay is determined with cross-correlation. The math for all that is in the ARTA manual.
ARTA Download
ARTA Download
Attachments
Hi,
This will have the effect that the equal phase contours of the wavefront are no longer plane (at infinity), but twisted. When signals enter ear canals from 2 stereo speakers, with the cross-talk, the ear signals have changed from the case where no "phase flip" would occur.
So, it would be a good idea to measure the whole wavefront (velosity vector) around the speaker. Who will do that? 😀
- Elias
This will have the effect that the equal phase contours of the wavefront are no longer plane (at infinity), but twisted. When signals enter ear canals from 2 stereo speakers, with the cross-talk, the ear signals have changed from the case where no "phase flip" would occur.
So, it would be a good idea to measure the whole wavefront (velosity vector) around the speaker. Who will do that? 😀
- Elias
Last edited:
SUM, I think it's fairly easy to know what CD or CBT people are talking about just by context. We do this with speech all the time-their, there, they're; it's, its; sew, so; Tso, TSOE etc... We do it in writing with thing like "bow", "mouth" or "bark" etc...
It seems people just want you to bring evidence for your claims. There's a wealth of info that doesn't agree with you claim. I don't think you'll find the participants in this thread to be the mainstream in audio or close minded, but we do many of us do believe the mainstream science of audio/acoustics and what our experience and experiments tell us works. Many of us will not believe that which contradicts the known science and personal experience and we have no evidence for or we can't produce any evidence of.
As much as I don't like admitting it(only b/c it seems I'm just parroting Dr. Geddes--which I'm not. My experiments just happen to agree with his conclusions), so far the things that have proved most telling to me are the polar response and that crazy impulse thing. I actually didn't believe him at the start of this thread, but recently after doing some CSDs and waterfalls, I now know what he was saying though I'm not sure I could put it to words. Still, I like looking at the spectrograms and the waterfalls, but they just shed a new light on old information. I would also add that knowing the CSD without the impulse would be useless to me.
Dan
It seems people just want you to bring evidence for your claims. There's a wealth of info that doesn't agree with you claim. I don't think you'll find the participants in this thread to be the mainstream in audio or close minded, but we do many of us do believe the mainstream science of audio/acoustics and what our experience and experiments tell us works. Many of us will not believe that which contradicts the known science and personal experience and we have no evidence for or we can't produce any evidence of.
As much as I don't like admitting it(only b/c it seems I'm just parroting Dr. Geddes--which I'm not. My experiments just happen to agree with his conclusions), so far the things that have proved most telling to me are the polar response and that crazy impulse thing. I actually didn't believe him at the start of this thread, but recently after doing some CSDs and waterfalls, I now know what he was saying though I'm not sure I could put it to words. Still, I like looking at the spectrograms and the waterfalls, but they just shed a new light on old information. I would also add that knowing the CSD without the impulse would be useless to me.
Dan
Last edited:
I just had a read through the Iasys manual and I think that is where a lot of Sum's misinformation comes from. Its flowery but vague descriptions of technical topics is written for the non-technical sound installer. I'm not totally sure what the coherence test is but that is where the "flipping" notion appears to come from. It is phase related but not phase, and might have given Sum the notion that phase rotation is a bad thing. The manual also describes dubious terms like the driver's "energy center".
Loudspeaker measurements is a complex topic and a company trying to simplify it for a not-technical user is admirable. But obscuring the science rather than explaining it is not the way to go.
David
Loudspeaker measurements is a complex topic and a company trying to simplify it for a not-technical user is admirable. But obscuring the science rather than explaining it is not the way to go.
David
It's quite understandable from that aspect. The problem arose when the system and its flowery but vague descriptions were used as the basis for declaring all other systems deficient. Worse yet, it appears that this was without understanding of the systems being essentially trashed and considered a "fad", from the release of MLSSA in 1987 through today. Possibly the lengthiest "fad" in history.🙄
Dave
Dave
Especially funny with ARTA since it can do IASYS-style or MLSSA-style measurements, all using the same math coded by the same guy. I haven't tried it yet but FR2 mode and ungated IMP mode should give the same FR and phase response as long as the same time-of-flight delay is set. The only difference is FR2 mode determines the delay automatically and IMP mode needs manual intervention.
Thanks John - excellent paper. Everyone should read it. Tackfully stated as well. Some people deserve tack, others do not.
Agreed, good paper.
FWIW, here are a few notes pulled from the ARTA manual. ARTA lets you choose MLS, swept sine, periodic noise or external signals for recording the impulse response. Short version, use swept sine if it's quiet, otherwise periodic pink noise and average several of them.
FWIW, here are a few notes pulled from the ARTA manual. ARTA lets you choose MLS, swept sine, periodic noise or external signals for recording the impulse response. Short version, use swept sine if it's quiet, otherwise periodic pink noise and average several of them.
Thats not quite accurate - there is an elaborate AES paper on this, I just don't rember the issue or names - in all cases, swept sine is prefered as it has the highest SNR (term used loosely here). Log swept sine has one other advantage and that is the time synchonicity of the nonlinearities - a big advantage.
I should also point out from John's comments that the use of a simple division of spectral functions is not the way most of the better measurements work. They use correlation, which is much more powerful. In the limit where the system is linear and noise free they all yield the same results, but for a noisy, nonlinear system log-swept sine is going to yield the best results possible. HOLM proves this better than I could ever describe it.
Good reading material. Thanks.
I'd like to add that measurement methods also depend on what the intention of the measurement is. Some types of measurements/signals do better in revealing specific types of problems than others; some present results that look better than others. Additionally, even with the same sampled data, different math processing techinques will yield result suitable for specific purposes. I have come across several cases where two software claim to use the same methods, but there will be some extreme cases where the results are quite different. This is where the programming techniques come into play, and only the top programmers really can tell you what's going on. Those type of guys are really hard to come by.
Whilst I simplified the description somewhat for the purposes of the discussion, I would point out that frequency domain multiplications and divisions are correlations 🙂
John
This comment sent me back to my texts. Correlation really requires concepts in "spectral density" functions and "expected values" and simple multiplications and divisions don't yield those frunctions or variables. In any case its all too obscure to be of much importance here, so I accept your point.
I should have elaborated. If we want to obtain the cross-correlation of a pair of time sequences an efficient method is to take the FFT of each, do an element-wise multiplication of one FFT by the complex conjugate of the other, then do an inverse FFT of the result which yields the cross-correlation. To see the kinship with FFT division, recall that complex division is the same as multiplication by the complex conjugate and scaling by the reciprocal of the squared magnitude. When using correlation that magnitude correction step typically comes later in the process, but the results are the same.
I thought about this some more, and I'm not in complete agreement now. The fundamental definitions of correlation involve limiting processes and can only be estimated based on finite data or a finite number of data sets. A single data set, as you are describing, can be used to "estimate" the correlation coefficient, but it is a very poor estimate. If the system is noise free and completely linear then "any" estimate is a good estimate, but if there is noise or non-linearities then this estimate gets bad very quickly and only more and more data sets can be used to improve the estimate. It can only be exact when an infinite number of data sets are used.
You probably knew all of this, but from your initial, admittedly simplistic, discussion it seemed like you were saying that a single time series, when divided would yield a "correlation", which I was interpreting as as "good" correlation.
For example, the coherence function for a single data set will always be unity - implying a perfect correlation between the input and the output, when this need not be the case at all. It is only after several data sets have been analyzed that the true cohernce will be seen and the true correlation of the input and the output can be noted.
Another example would be where there was hum in the output spectra, but not in the input spectra. A single data set division as you described would associate this "noise" as part of the transfer function - which it is not. Only after several data sets were analyzed would the coherence at that frequency fall indicating that the "estimate" of the transfer function at that frequency was highly uncertain. The initial correlation at that frequency would be highly erronious.
Way too far off topic here, and probably only the two of us would care about any of this anyways.
Here's a question for you guys. When I do FR graphs with REW, I always test the first graph using a long sweep, and the repeat it with a quicker sweep. They always look the same, but I know the long sweep has a higher SNR. Is there any practical advantage to using the long sweep? None that I know of, but I bet one of you guys could help me with this. There must be a reason for it in the programming that is probably obvious to most of you.
Thanks,
Dan
Thanks,
Dan
for what it's worth, I use praxis. I can run sweeps, mls, impulse, these can be weighted and run in a couple ways. All of them are at least internally consistent. Additionally, I use the measured phase. I don't assume minimum phase behavior, though, like everyone else here I believe most drivers under most conditions are minimum phase.
My measured phase data when placed into the usual modeling software (i.e. lspCAD or SE) yields essentially the same result whether using measured phase or assuming minimum phase and letting the software extrapolate the phase at a particular point in space.
While I've seen some dubious measurement data posted now and then on the net, most of it comes from very short measurement intervals-sloppy indoor measurements in small rooms etc. In general the measurements from holm/rew/clio/praxis/SE and similar all seem to generate respectably accurate measurements for phase as well as FR. No assumptions about minimum phase behavior are necessary. You can design with measured phase, or generate a driver model with minimum phase if you wish. They both yield the same result. As they should.
- Status
- Not open for further replies.