More on the world above 20 KHz - a reference

[Joe Gwinn]

It occurred to me that David Blackmer had most likely published data backing his theories up, so I did some googling.

Read the third and following paragraphs of the following: Watermarking, High-Resolution Audio Big Topics at AES | Stereophile.com

There must be a proceeding or the like, which I'll try to obtain.


The obit is here: http://www.stereophile.com/news/11303/

[Mike Sulzer]

It occurred to me that David Blackmer had most likely published data backing his theories up, so I did some googling.

Read the third and following paragraphs of the following: Watermarking, High-Resolution Audio Big Topics at AES | Stereophile.com

There must be a proceeding or the like, which I'll try to obtain.


The obit is here: dbx Founder David Blackmer Dies | Stereophile.com

A fascinating discussion ensued in which an alternate model of human hearing was presented as a possible explanation for the reason high-resolution audio sounds better. Instead of simply detecting tones, the hearing system might also detect impulses, or "clicks"—localization cues that arrive at the ears within a 10-microsecond window.

Strawman warning! No one who understands anything said the ear detects only tones, but rather that it analyzes the output in time of a set of "filters". 10 microseconds! There must be some path around the resonant membrane straight into the nerves!

[Joe Gwinn]

Strawman warning! No one who understands anything said the ear detects only tones, but rather that it analyzes the output in time of a set of "filters". 10 microseconds! There must be some path around the resonant membrane straight into the nerves!

Yep.

[Mike Sulzer]

Yep.

But it is the nerves that cannot respond so fast. That is the purpose of the resonant membrane with the many resonant sections; something has to respond fast enough to detect the vibrations.

[Joe Gwinn]

But it is the nerves that cannot respond so fast. That is the purpose of the resonant membrane with the many resonant sections; something has to respond fast enough to detect the vibrations.

Not necessarily. It isn't necessary that individual nerve cells be that fast. Based on countless experiments, nervous systems use population codes, where it's the average over thousands of cells that matters. Think of it as a form of dithering.

[Joe Gwinn]

By googling on the names of some of the people listed as participants in the AES discussion cited in posting one of the present thread, I found some good references into the issues. The key was the unexpected observation that 96 KBit sampling sounded more transparent than 44.1, and 192 more than 96. The question was why.

Green in particular has a good set of references.

It turns out that the brickwall anti-aliasing filters are a big part of the problem. Note that Bessel filters are an approximation to Gaussian.

[Mike Sulzer]

By googling on the names of some of the people listed as participants in the AES discussion cited in posting one of the present thread, I found some good references into the issues. The key was the unexpected observation that 96 KBit sampling sounded more transparent than 44.1, and 192 more than 96. The question was why.

Green in particular has a good set of references.

It turns out that the brickwall anti-aliasing filters are a big part of the problem. Note that Bessel filters are an approximation to Gaussian.

These papers say that it is the effect of the processing on the frequencies below 20 KHz that matters, that is, sample faster because it makes the processing better where it matters. Or in other words, there is no world above 20 KHz. That makes a lot of sense to me. Some of it is just speculation, of course.

That said, there are some serious misconceptions in the Story paper, right from the beginning. His claim that an impulse and a sample of random noise have the same spectrum is wrong. It is only the process that has the flat spectrum; any single sample has a jagged spectrum. What he might have meant is if you take an impulse and stretch it with phase distortion, you get something that sounds different even though the spectrum is the same.


Then there is this:

One can get oneself into a bit of a twist thinking about the energy in the ringing. After all, if it is in the audio band, allowing extra energy at higher frequencies through the system surely cannot cancel out some that is in the audio band? It does, though - so although we may not be able to hear energy above 20 kHz, its presence is mathematically necessary to localise the energy in signals below 20 kHz, and it is possible (and our contention) that we can hear its absence in signals with substantial high frequency content. A high sample rate system allows it through (fact) - and allows the high frequency signals to sound more natural (contention) but allowing better spatial energy localisation (fact).

That is just totally wrong. It is hard for me to believe that someone with an engineering or scientific education would write that.

[Joe Gwinn]

These papers say that it is the effect of the processing on the frequencies below 20 KHz that matters, that is, sample faster because it makes the processing better where it matters. Or in other words, there is no world above 20 KHz. That makes a lot of sense to me.

The papers did say that a big effect was the 20 KHz brickwall filter, and that a higher sample rate allowed one to use a more gradual filter, the ideal being Gaussian, this relaxation reducing ringing and smearing of the signal below 20 KHz.

But they also said that one can detect 10 microsecond differences in arrival time at the two ears, and if this was messed up, transparency was degraded by the inconsistency. Each ear by itself is far less sensitive to time delay effects, more like a fraction of a millisecond. (The discussion of inter and intra time delay differences is hard to follow. I had to read it three times.)

Some of it is just speculation, of course.

Well, yes. This is a research frontier. Progress depends on new ideas. Most of the new ideas are ultimately refuted, but without a steady supply of new ideas, progress stops.

That said, there are some serious misconceptions in the Story paper, right from the beginning. His claim that an impulse and a sample of random noise have the same spectrum is wrong. It is only the process that has the flat spectrum; any single sample has a jagged spectrum. What he might have meant is if you take an impulse and stretch it with phase distortion, you get something that sounds different even though the spectrum is the same.

Umm. Story was drawing attention to the fact that an impulse and white noise have the same power spectrum in the limit - totally flat. The only remaining degree of freedom is phase.

Another way to approach this is to consider the autocorrelation function of gaussian noise, and of an impulse - in the limit, they are the same. One standard way to compute autocorrelation consists of taking the Fourier Transform, computing the power spectrum (thus losing all phase information), and taking the Inverse Fourier Transform to yield the autocorrelation function (of offset time).

That is just totally wrong. It is hard for me to believe that someone with an engineering or scientific education would write that.

Well, this is a research frontier, and things are by definition not yet settled. Nor do we yet know if the quoted statement is right or wrong. And the ad hominem comment is not helpful. One must instead show exactly why the statement is wrong.

[Mike Sulzer]

I agree that times relating to differences between signals in two ears could be faster than for one. My experience is that 10 micro seconds is much faster than I can hear! But my hearing is down in one ear. Perhaps others can do much better. I will post some functions later in the week.

You do not hear limits; you hear specific signals. If you make a sample of random noise, it does not have a flat spectrum. All frequencies are there at some level (zero probability of zero level), but the impulse really does have them all at equal levels (limited by the filter, of course).

When you listen to samples of random noise, each one has a different spectrum. The ear/brain has a way of handling random signals, and certainly some averaging is involved at some level; for example, white and pink noise sound different. Certainly what the ear/brain does with an impulse and with random noise are two very different things. The two do not make a correct example of how the ear/brain arrives at different answers even though the spectra are the same. They are not the same in the same sense.

Nor do we yet know if the quoted statement is right or wrong.

Of course it is wrong. Allowing in more energy does not cancel out energy, as he says ("It does, though"). The filtered signal has less energy.

It does, though - so although we may not be able to hear energy above 20 kHz, its presence is mathematically necessary to localise the energy in signals below 20 kHz

Energy in one frequency range does not localize energy in another range. Take out the high frequencies and you have a new signal that is less localized. Put them back in, and you get the original signal back. That is all.

[Joe Gwinn]

I agree that times relating to differences between signals in two ears could be faster than for one. My experience is that 10 micro seconds is much faster than I can hear! But my hearing is down in one ear. Perhaps others can do much better. I will post some functions later in the week.

You do not hear limits; you hear specific signals. If you make a sample of random noise, it does not have a flat spectrum. All frequencies are there at some level (zero probability of zero level), but the impulse really does have them all at equal levels (limited by the filter, of course).

When you listen to samples of random noise, each one has a different spectrum. The ear/brain has a way of handling random signals, and certainly some averaging is involved at some level; for example, white and pink noise sound different. Certainly what the ear/brain does with an impulse and with random noise are two very different things. The two do not make a correct example of how the ear/brain arrives at different answers even though the spectra are the same. They are not the same in the same sense.

One specific kind of random noise was specified in the impulse example, Gaussian, which by definition has a flat power spectrum.

Of course it is wrong. Allowing in more energy does not cancel out energy, as he says ("It does, though"). The filtered signal has less energy.
Energy in one frequency range does not localize energy in another range. Take out the high frequencies and you have a new signal that is less localized. Put them back in, and you get the original signal back. That is all.

You're misreading it, and missing the author's point. He is refuting is the no doubt common argument that all that interesting stuff that happens above 20 KHz cannot affect what happens below 20 KHz. It is the mathematical properties of filters (and fourier transforms) that force one to consider the realm above 20 KHz.

And always remember that these efforts arose from the attempt to explain an unexpected observation, that raising the sample rate to 192 KHz greatly improved transparency.

[Mike Sulzer]

One specific kind of random noise was specified in the impulse example, Gaussian, which by definition has a flat power spectrum.

The meaning of the statement in bold is that the random process has a flat spectrum. Any sample of that process, no matter how long, does not (except the degenerate case where all values but one are zero). He compares two waveforms in figure 1. He says they have the same spectrum. They do not.

You're misreading it, and missing the author's point. He is refuting is the no doubt common argument that all that interesting stuff that happens above 20 KHz cannot affect what happens below 20 KHz. It is the mathematical properties of filters (and fourier transforms) that force one to consider the realm above 20 KHz.

And always remember that these efforts arose from the attempt to explain an unexpected observation, that raising the sample rate to 192 KHz greatly improved transparency.

Story:

After all, if it is in the audio band, allowing extra energy at higher frequencies through the system surely cannot cancel out some that is in the audio band? It does, though

Could you explain how you read that statement? I think it is clear that it is wrong.

In my previous post I made certain specific statements. Your answers are general statements that do not address the specifics of what I wrote. That can never be convincing.

[Mike Sulzer]

Perhaps an example would help. The attached figure shows a sample of Gaussian random noise (top) and the power spectrum of that waveform (bottom) ([/QUOTE]GauRanAndPS.pdf). The spectrum is not flat; it is not like that of an impulse or a band limited impulse.

Another example: suppose someone suggested using a sample of random noise instead of a Barker code or chirp for pulse compression on your new radar. One (of several) objections would be that the side lobe level from a typical sample of random noise is really quite bad. You could, of course, search through a great many samples and find one that was pretty good. It is the acf of the specific sample that matters, not the acf of the process.

One specific kind of random noise was specified in the impulse example, Gaussian, which by definition has a flat power spectrum.

[guitician]

I'm glad that the 1957 tweeters that I listen to every day were designed with a frequency response far beyond audibility.

[Joe Gwinn]

The meaning of the statement in bold is that the random process has a flat spectrum. Any sample of that process, no matter how long, does not (except the degenerate case where all values but one are zero). He compares two waveforms in figure 1. He says they have the same spectrum. They do not.

Sure they do, in the limit. Or, more commonly in the theory, the ensemble is flat. But in ordinary speech, one does not repeat these qualifications every time - they are understood.

Could you explain how you read that statement? I think it is clear that it is wrong.

Well, you need the whole quote to understand what is being said. Pulling stuff out of context makes it harder to understand. From posting #7:

One can get oneself into a bit of a twist thinking about the energy in the ringing. After all, if it is in the audio band, allowing extra energy at higher frequencies through the system surely cannot cancel out some that is in the audio band? It does, though - so although we may not be able to hear energy above 20 kHz, its presence is mathematically necessary to localise the energy in signals below 20 kHz, and it is possible (and our contention) that we can hear its absence in signals with substantial high frequency content. A high sample rate system allows it through (fact) - and allows the high frequency signals to sound more natural (contention) but allowing better spatial energy localisation (fact).

While Story isn't being mathematically precise, his meaning is clear, and his presentation is pitched correctly for his intended audience. Remember that Story's objective is to refute some common arguments heard from that audience, composed largely of audio engineers, and he is speaking their language.

In my previous post I made certain specific statements. Your answers are general statements that do not address the specifics of what I wrote. That can never be convincing.

Guilty. If I disagree with a fundamental premise of block of text, I address the fundamental premise first, and ignore everything depending on that premise.

After the matter of the premise is settled, then we can move on to such dependent statements as have survived.

[Joe Gwinn]

Perhaps an example would help. The attached figure shows a sample of Gaussian random noise (top) and the power spectrum of that waveform (bottom) ([ATTACH]27591[/ATTACH]). The spectrum is not flat; it is not like that of an impulse or a band limited impulse.

Another example: suppose someone suggested using a sample of random noise instead of a Barker code or chirp for pulse compression on your new radar. One (of several) objections would be that the side lobe level from a typical sample of random noise is really quite bad. You could, of course, search through a great many samples and find one that was pretty good. It is the acf of the specific sample that matters, not the acf of the process.

Flatness issue addressed in a parallel thread.

Fact is, we do use random sequences, so long as they are long enough. Barker codes are far too short for modern radars. Some people use Barker conds with complex elements, other people combine multiple orthogonal Barker codes in interesting ways, but even so, a million-sample pseudorandom sequence is near-optimal. (Simple shift-register sequences are not usually used for this.)

Gold codes are also used: Gold code - Wikipedia, the free encyclopedia

These are bigger than Barker codes, but still can be too short.

By the way, what does "acf" stand for?