The Sound of Science

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
amplifierguru said:
"Measurements and hi-end gear don't mix. "

Now that's PURE marketing. ...


Hmm... I think I've been misunderstood. I'll try to re-phrase it all.

I don't like over-enthusiastic salesmen. I'm a DIY'er myself, not a salesman. I have no interests in promoting certain brands. The only type I like to encourage is DIY.

What I meant was:
I think I know why the fellow didn't suceed in starting a thread based on science and measurements:
1) Few have the gear to measure it all (suficiently accurate, anyway).

2) Not all are INTERESTED in measurements. Numbers tend to take the woodoo out of sound reproduction.

Jennice
 
Disabled Account
Joined 2002
Jeez . . .
Science is science.
Audio is audio.

Tech specification and measurements are good details backing up when I am nodding towards the sound of a product. And, I could use them to provide good matching in the chain. Otherwise, they are just papers . . . can not persuade me . . .

Regards
jH
 
jh6you said:
... Otherwise, they are just papers . . . can not persuade me . . .

Regards
jH

Which is why I don't think a thjread could be started on that.
Indeed, one could start a thread to aim for the best MEASURING amp around (if enough participants have the equipment for it), but it is not likely to be the amp that people like the sound of.

Jennice - signing off...
 
About Anli's thread: he wanted to know which publications link specific distortions to audibility. It had nothing to do with "hi end gear doesn't like measurements". His question was about how to relate measurements to perception.

The question was a very worthy one, and I for once find it funny that really no one contributed. I didn't either, because all documents or links that I have, relate either to
- what humans can hear (regardless of audio)
- various distortions
- perception of distortion, but not relating a specific kind of distortion to audibility with Anli's required publicly available double blind listening test data.

The ABX people for instance have a lot of demo files where you can find out for yourself whwther you can hear a specific distortion. Very valuable. But no one seems to have compiled the data into some solid test.

I am sure JAES must have published review articles about audibility of different distortion mechanisms at some points in history. But I am not a member, and no member seemed to care either. OTOH Anli asked for publicly available info, and the scholarly journals are not free...

My conclusion is that it's not just the "audiophile" camp in this forum that failed to "prove" , by quoting specific scholarly test data, that some "nonstandard" distortions are *audible*. The "engineering" camp in this forum equally failed to "prove" by citations, that low level standard distortions, as claimed, are *inaudible*.

At least, none posted to Anli's thread.

BTW regarding Anli's thread the GedLee site comes to mind (demonstrating how some kind of distortions can be audible at say 0.01 % and others inaudible at say 3%), and the Cheever thesis. But again both fail to fulfill Anli's standards: while making good points they both are short of data from well done double blind tests.

Maybe a good idea for the forum: next time someone says "we can't hear 0.05% THD" or "we can hear the typical solid state 0.001% 7th harmonic very well" they get sinbinned unless they add a full citation. Like "according to Schwarzensteinrichsen et al, 'On the Nature of the audible Universe' 1899 pp59" :smash: .

How about that? THAT would be science.
 
MBK,

Thanks for your kind acceptance! :)

_________________________

To emphasize the situation we have, such analogy may be done: if we have problems with hearing, we go to audiology specialist. She produces different measurements. Last ones help she to carry her responsiblity for our helth.

When we want to listen to a music, we go to by an amp. But sellers have not any responsiblity for our hearing, so, (continue youself... :))


Andrew aka anli
 
How about that? THAT would be science.

Well it would be a damn good start.

Gedlee and Cheever are both very interesting, but as you say, both somewhat flawed. Cheever is entirely a theoretical work, and it stands damned due to its central precept having no provenance. Which is very sad.

The issue of distortion claims are a great mix of problems, and the need to exactly specify the history of the claim, and even more importantly, understand the precise limits of the claim are a problem that has beset audio for decades.

So many of these limits to audibility claims and tests are based upon a limited ideal of the nature of the signal. For a start most only talk in terms of THD+N. So no time varying products, and as GedLee points out, no understanding of the nature of sound perception mechanisms in the ear or brain. Indeed the use of THD of itself is the result of another flaw in thinking - that of using harmonic analysis - and hence the (Fast) Fourier transform as the basis of all analysis. There are fundamental flaws in using a FFT that are commonly not understood. Interestingly they were understood by Tukey himself, but lost as the default method became the accepted wisdom of the right way to do things.

There is a great issue of hammer and nails (when the only tool you have is a hammer, everything looks like a nail.) If your default tool is an FFT, everything looks like a harmonic series, and the error looks like THD+N. And the tools oblige by presenting everything in this light. But the use of Fourier analysis requires assumptions that are actually violated by music. Violated in the small print, but violated none-the-less. It seems very likey that an approach based upon maxium entropy methods with a basis function derived from some form of active resonator (working in the same manner as the ear) might be the right answer - or at least a better answer.
 
Francis,

you have the credentials - I'd be very interested to hear a short expose on the shortcomings of FFT.

AFAIK the fundamental flaw is that FFT presupposes a *periodic* signal, which music is not. Since music is not periodic, wave form = time domain info, matters, which FFT ignores. Now, this is/was my understanding. BTW this also relates to all claims of "delay" vs. "phase" - IMO "phase" as a mathematical construct only makes sense in the context of a periodical signal - for an impulse, I'd say the use of the term "delay" should be allowed ;-) unlike many others here.

But I am not a physicist nor EE thus I can't claim this with authority - which you could provide ;-). So, what is your take on FFT?
 
MBK, two things:

First, if you eliminate JAES and JASA because they're not available free on-line, you might as well give up. That's where all the real work gets published.

The notion about FFT and periodicity is a flawed one- you can make anything periodic without changing it in the least. Consider an arbitrary function, f(t). It starts at some time (which can call t = 0) and ends at some time to. Maybe it's an orchestra playing a movement from Dvorak's Ninth, maybe it's the sound of a glass breaking, whatever. Now, we make a reasonable assumption- that we can displace the function in time without altering its properties, i.e., there is a time displacement symmetry. That means that if we record this signal with a perfect recorder, we can play it back unaltered.

You can construct a periodic function f1(t) by translating f(t) forward by to. IOW, you could play the tape once, then play it again. And again. And again. You then have a function which is equal to the sum of the original function, the original function displaced by to, the original function displaced by 2to, the original function displaced by 3to.... It is apparent that between any two starting and stopping points, the new function is periodic with period to, and that between any points nto and (n+1)to, with n being an integer, it is identical to the starting function.
 
These kind of discussions are great, but it all really needs to be put into perspective.

How often are our sources well recorded stereo?(i.e stereo pair, not pan potted mono mic'd), very rarely(unless of course you're a classical fan, but even here, there's no promises)

How often are our recordings not eq'd, compressed etc?

My point being that even with an supposedly *accurate* reproduction system, all we are getting is accuracy to what's on the disc. How often is what's on the disc *accurate* to what was being performed/recorded?

Even with a stereo pair, it's still not going to sound *life like*(I've discussed this elsewhere, I'll refrain from bringing it up here)

So I say let people have their valve amplifiers, their turntables and full range drivers. Yes, there are more *accurate* alternatives available, but with the likelyhood of the recording being *accurate* pretty low, can't we tweak it further to suit our tastes/needs/inclination at the time?.

Also, another thing that rarely gets brought up, is that some kinds of harmonic distortion can sound *nice*. A lot of people think of distortion as purely being some nasty artifact to be avoided at all costs, this is not always true.

There is no reason not to try to attain greater *accuracy* of reproduction, this is also very amicable.

But think about this, faced with the perfect reproduction system, that sounded as close to the actual *sound* being produced at the time of recording, how many of us would actually be happy with it?

I dare say there would be a good amount who would not be....
 
SY,

First, if you eliminate JAES and JASA because they're not available free on-line, you might as well give up. That's where all the real work gets published.

I agree - but this is what Anli had in mind in his thread, a collection of publicly available data. IMO it would be worthy to start a Wiki with a table containing info on distortion perception of varying quality:
- sound bits sites a la ABX, to test for yourself
- magazine articles, qualitative, and essays
- semi scholarly tests
- hardcore , actual scholarly tests, including those you need to buy. For those ones, ppl who own the articles could either post the abstract , if permissible, or collect the main points themselves and summarize them. This should qualify as a citation and thus not require copyright clearance.

The symphony example: we've had a similar discussion before... I do not deny that you *can* do an FFT of a symphony, treating it as a single cycle of a hypothetically periodic signal. But wouldn't you agree that the meaning of the symphony is almost entirely in the time domain, and an FFT of a symphony completely meaningless? Now extrapolate this down to a short transient within this symphony: its time domain form matters! And the FFT tells us nothing about that.

Thought experiment:

You take the symphony and replay it, removing an instrument. The musical message remains the same.

You take the symphony and replace the entire orchestra by a single piano. The musical message remains quite comparable.

In both cases the "distortion" is drastic on an FFT (both linear and non linear).

Now you take the original symphony recording , orchestra and all, and randomize the waveform but using *exactly* the same original symphony's FFT data. You get pink noise. Now what?

The FFT does *not* tell us something *essential* about the musical info, because the musical info is contained not in the *periodicity* of the music's elements, but in their time domain distribution, a very specific distribution.

So while you may *mimic* periodicity as an assumption permissible in order to be able to perform an FFT, you entirely miss the point which lies in the non periodicity of the original signal.

I don't say that the math is wrong. You *can* do an FFT of a transient. But it is almost entirely irrelevant to what we want to know.

IMO.
 
OK, the flaw in using FFT is exactly the issue of periodicity. Actually you can't simply take the sample set and repeat it ad-infinitum - there is the small issue of the join. Where the samples begin and end - or are in principle joined to make the infinite periodic sample - must be smooth. If they are not the join introduces artefacts into the system. Essentially this is ringing. Smooth is variously defined, and interestingly a Fourier series itself can be used as a definition of smooth.

Anyway, in order to avoid the problems with the ends of the sample the sample must be windowed. This is simply that act of scaling the samples so that those samples at the ends are scaled to close to zero, and those in the middle of the sample space retain their full value. The window function needs to be smooth of itself avoid introducing further artefacts (although not all are). There is nothing special about these window functions, they have names like Blackman, Hanning, Hamming, but are really nothing more than using well know functions (like cosine, cosine squared, trapezoid) to provide a smooth window.

The problem with windowing is that it intrinsically throws away some data. But it throws away data in a manner that allows the FFT to more easily match the remaining data to a harmonic series. In a manner of speaking it throws away the data that conflicts with the premise that the data is all harmonic in structure (a bit of an oversimplification here.)
Indeed the standard windowing techniques actually damage the ability of the FFT to resolve harmonic structure too - but the manner in which this happens has become so entrenched in standard practice that most users think this limit to resolution of the FFT is intrinsic - which it isn't - it turns out to be much finer.

FFt is simply a rather useful implementation of the DFT - the Discrete Fourier Transform. The DFT is simply a model. It is always true that if you take a sample set, take the DFT you have exactly captured the nature of the sample set in frequency space, indeed (crucially) you can reverse the process and get the exact sample set back again (baring rounding errors) - interesting the inverse of the DFT is the DFT (along with a re-normalisation of the data.) There are an infinite number of equivalent transforms to the DFT. Discrete Cosine Transform (famous in Jpeg compression), Hadamard Transform (used in maximal length sequence audio analysers) are others.

The point is not that you can transform the data from one space to another - that is true. The problem is what it means in that other space. Taking an FFT of an audio signal nicely shows us the harmonically related information. But it hides the enharmonically correlated information as noise. And worse, the windowing function has intrinsically reduced the impact of this information, resulting in an artificially low noise floor that incorrectly suggests that the FFT has captured more information about the sound than it really has.
 
The effect of the join goes to zero as the period gets large. As you hinted, even that error can be greatly reduced by an appropriate choice of window. The effects of various windows on the resulting spectra are very well-known and characterized.

The idea that an FFT gives you every piece of information you could possibly want is a straw man. But it's not because of some "problem" with periodicity.
 
One should be aware that only the transformation of the time-domain into the frequency-domain gives an unambiguous result but not vice-versa if one transforms large sequences of an acoustical event.
If we now take into account that it is rather unlikely that even one period of an instrument (let alone a full orchestra) is the same as the one following it we could question the validity of measurements that only look at the (static) frequency response (and THD etc) when it comes to evaluating the performance of audio equipment.


Regards

Charles
 
It isn't intended as a straw man. FFT analysis is very powerful, we all agree about that. The issue is the manner in which it has been applied as essentially the only measure of important distortion in audio. The point I'm trying to make is that if missapplied it becomes a self fulfilling prophesy. Windowing is more black art than science. Typically a compromise between spectral resolution and ringing. If you know what it is you are looking for this isn't a problem.

However there is the traditional school of thought in audio that says "you can't hear distortion below x%" and proceeds to use FFT analysis to both measure and bolster the opinion. But windowing the FFT remains a damaging operation. If the argument is about thus far unmeasured artifacts, the residual information lost in the windowing operation becomes important. The information is lost - that is intrinsic. Back on thread topic - it is this subtle questioning of the assumptions that underpin the models used that keeps the science honest.

The "problem" with periodicity is simply that it is an axiom of the technique that has interesting and sometimes overlooked implications.
 
Francis,

thanks, very informative, I wasn't even thinking of those kinds of problems.

SY,

I didn't mean to imply anyone said FFT gives all the information. But just ignoring the time domain and effects therein isn't going to help either. Quoted specs for amps are invariably in the frequency domain.

And regarding
The effect of the join goes to zero as the period gets large.
etc

a lot has come from small problems that seemed too insignificant to mention, at first. Jitter for instance.

Another thing that puzzles me: I recently browsed the Class D forums. Lo and behold, here we have ppl who claim excruciating transparency from amps that are a) not a baby of the bid magazine reviewers, b) inexpensive and c) have been shunned in high end circles for a long time. Now ppl claim they are exceptional. THD seems quite low but not exceedingly so on average. One of the THD plots I've seen in those threads even shows a perfect series of harmonics at -80 dB ad infinitum... still thay are loved by their fans. Another puzzle. What's going on?
 
Cutting-edge research

A little-known fact that affects sound quality more than previously recognised is that a schematic drawn on a Windows system has markedly poorer performance than the same one drawn on a Linux system. This only applies to solid-state circuits. Valve circuits perform best when drawn with a pencil and paper. It's true... :Pinoc: :Pinoc: :Pinoc:
 
I think it should be obvious- fashion and sonics are not the same thing.

SY,

are you playing sort of an act here? Because you can't be serious in simply dismissing *any* occurrence of opinion outside your approved worldview as fashion, unscientific drivel, illusion, and the like. The graveyard of science is full of that kind of statements.

As to

There's a 1:1 correspondence between a time domain function and its Fourier transform. And vice versa.

If I play a 1 kHz tone for 1 second, then a 2 kHz tone for 1 second, and I do an FFT, I get a 1 kHz band and a 2 kHz band.
If I play the 2 kHz tone first, and then the 1 kHz tone, and I do an FFT, I get ... a 1 kHz band and a 2 kHz band.

Does not look unambiguous to me.

Nevermind the windowing artefacts.

BTW each and every book I open that mentions FFT warns of the possible pitfalls. Why else would there be so many ways to do an FFT? Almost as many as amplifier topologies :clown:
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.