Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: "Audiophile" listening event @ Definitive Audio in Seattle (Read 153992 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #350 –
Quote from: Notat on
I think it is safe conclude there's information in the LSBs if we hear a correlated signal there. I'm not convinced the converse is true.
I agree. (I've got to stop doing that!).

This is actually quite complicated. Hence my interest in what on the surface seemed quite a simple statement from JA.

It's this complicated...
http://www.hydrogenaudio.org/forums/index....st&p=746571

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #351 –
Quote from: 2Bdecided on
And what you see depends largely on the signal content and the settings of the FFT (or similar).

You can use a longer FFT to "see" a pure tone further into the noise.

The ear's auditory filters cannot be adjusted in the same way .

In other words, it's quite possible to see something which is entirely inaudible - the apparently "lower" noise floor actually masks the "higher level" tone within the auditory filter in the ear.

Cheers,
David.


The ear's maximum integration time is circa 200 milliseconds, or perhaps less, but more importantly, the bandwidth of the ear's filter is what really matters here, and they are quite broad, compared to an FFT, so in fact it is easy to develop a signal which stands out like a lightbulb in an FFT which is totally inaudable under any circumstances above the masking noise.

In an FFT, you're plotting not energy per root Hz, but rather energy per BIN. All else follows, when you realize that noise gain is RMS, while signal gain is amplitude sum.
-----
J. D. (jj) Johnston
 

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #352 –
I guess no one is going to present any real world 24-bit recordings with non-noise-like 8 LSBs?

I guess no one is going to take a stab at defining what is/isn't "signal related information" in those last 8 LSBs either?

You're all cowards

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #353 –
Quote from: 2Bdecided on
I guess no one is going to present any real world 24-bit recordings with non-noise-like 8 LSBs?

I guess no one is going to take a stab at defining what is/isn't "signal related information" in those last 8 LSBs either?

You're all cowards


Nahh, I showed that the presnce/non presence of signal related information in the LSBs is not the important thing.

The capability to transmit high resolution information of a digital signal is not just in the LSBs. The capability to transmit high resolution information is in the uniform spacing between *all* of the different signal levels that it represents. The MSBs are just as important. If you are building a DAC the LSBs have to be more accurate than the LSBs by far.

The idea that the LSBs are the unique home of low level signals is yet another audiophile myth.

The noise floor of sound in the natural world is set by the Brownian Motion of air.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #354 –
You're still over thinking it.

JA claimed signal-correlated activity in those last 8 LSBs.

I'd like to see a recording that shows this. And maybe consider how one decides what's signal correlated or not.

I really don't want another lecture on why I shouldn't be interested in this.

EDIT: Sorry to sound harsh, but you keep jumping in and answering a different question, and talking to me like a grandparent would talk to their three year old. It's getting really tiring.

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #355 –
Correlated to what?

1. Correlated to the analog input fed into the ADC? That's easy, every bit is correlated within the full extend an ADC's capabilities, i.e. the complete word length minus the ADC's noise, which is measurable. You'll find at least 120 dB worth of undoubtedly correlated content within the digital output of a good ADC.

2. Correlated to the acoustic phenomenon before any (analog) electronic capture? That's not as easy, but totally really unrelated to the MSB/LSB juggling, that has been going on here lately. The question remains the same, whether the result is digitally captured or not.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #356 –
Quote from: googlebot on
Correlated to what?

1. Correlated to the analog input fed into the ADC? That's easy, every bit is correlated within the full extend an ADC's capabilities, i.e. the complete word length minus the ADC's noise, which is measurable. You'll find at least 120 dB worth of undoubtedly correlated content within the digital output of a good ADC.

Firstly, let me say that I think you're hitting the point now. There was a statement that something was signal-correlated, without (AFAICT) any real details about what that means, or how we could prove/disprove it. So, yes, this is exactly my point. The statement is ill defined and difficult to prove either way.

But running with your question, I think what you've said above is more or less true. But that begs another question: are the bits below the ADC's noise floor "correlated" to what was fed into it?

What about the noise due to other electronics in the chain? Are the bits below this noise floor correlated to the signal that came out of the microphone (before other noise was added)?

What about the noise in the room at the point where the microphones were placed (or at least any white-noise-like component of it)? Are the bits below this noise floor correlated to the sound that was in the room a bit closer to the instruments themselves (i.e. where the SNR was a little higher?).

I think, in the terms WernerO put it...
Quote from: WernerO on
Quote from: Notat on
There almost always is useful information below the noise floor.


Below the summed/integrated noise: yes.

Below the spectrally-local noise density floor: no, not really. Try listening to a fade to
noise while monitoring it on a decent real time spectrometer.
...the correct answer is no. While it's true that you can twiddle the spectrum analysis to pull out this or that, there's an uncertainty-principle time/frequency relationship beyond which you cannot pull something out of noise. Beyond this, the thing isn't there. Or if it is, there's no way of determining whether it really is or isn't, so basically it isn't!


So I propose there is a level below the noise at which things below the noise really are lost, if you accept some finite limit on integration time / FFT length.

This means that, for a given spectral noise level, there are only so many bits you need to store all the real data that's in the signal. If you use any more bits, those extra bits aren't storing anything that is in any way useful.


I could be wrong of course, but that's the way I see it. And from this it follows on that a given recording only ever needs X bits. You don't even need to consider replay level, human ears, masking, or even dither. The only caveat is that you'll raise the noise floor by a predictable amount (quantisation noise). If you don't want to do that, don't quantise - and for goodness sake don't process the signal or put it in the analogue domain .


So, back to the original claim. Are there 24-bit recordings where those last 8-bits contain something above the spectral noise floor, hence signal-correlated, hence at least potentially useful? I'm sure there must be, but I'm still waiting for a pointer to one.

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #357 –
Quote from: 2Bdecided on
You're still over thinking it.


Nope.

I see a lot of people trying to answer an audiophile myth. Dispel the myth if you want the answer you seek.

Quote
JA claimed signal-correlated activity in those last 8 LSBs.

I'd like to see a recording that shows this.


First state the problem in a reasonable way. This low order bit thing is a red herring!  It is one of JA's audiophile myths,

If low order bits were so important  to reproduciing small details, DAC monotonicity from rail to rail would not have been such a concern back in the days when it wasn't inherent.

Here's a good rule: Bits are not nearly as importeant as integrity of the data that they represent. It should be a short step from there to a complete dropping of discussion of just low order bits.


Quote
And maybe consider how one decides what's signal correlated or not.


Quote
I really don't want another lecture on why I shouldn't be interested in this.


You seem to be edging towards how things are, here:

Quote
1. Correlated to the analog input fed into the ADC? That's easy, every bit is correlated within the full extend an ADC's capabilities, i.e. the complete word length minus the ADC's noise, which is measurable. You'll find at least 120 dB worth of undoubtedly correlated content within the digital output of a good ADC.


EDIT: Sorry to sound harsh, but you keep jumping in and answering a different question, and talking to me like a grandparent would talk to their three year old. It's getting really tiring.


In my view you keep throwing the cereal bowl onto the floor. ;-)

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #358 –
Quote from: Arnold B. Krueger on
Quote from: 2Bdecided on
You're still over thinking it.


Nope.

I see a lot of people trying to answer an audiophile myth. Dispel the myth if you want the answer you seek.

Quote
JA claimed signal-correlated activity in those last 8 LSBs.

I'd like to see a recording that shows this.


First state the problem in a reasonable way. This low order bit thing is a red herring!  It is one of JA's audiophile myths

Whether it's a red herring or not depends on what you are trying to find out.

Signal correlated content in the LSBs is certainly not a "myth" - for a given recording, it can only be:
a) demonstrably true, or
b) not demonstrably true.


A claim that it has some relevance to what we hear, or an implication that it proves 16-bits aren't enough, or a suggestion that this signal is the difference between a 24-bit master and a correctly dithered 16-bit version - those would be audiophile myths. Those are unsubstantiated claims and/or just plain wrong.

But I am not going there, neither am I intending to go there.

So please, for the last time, and I'm asking nicely here - please stop replying to this question with answers to different questions that I haven't asked. Thank you.

Cheers,
David.

P.S. If you have time to waste, please go and download lossyWAV and listen to what it takes away. Disable noise shaping, or grab an older version without it.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #359 –
Quote from: 2Bdecided on
I see a lot of people trying to answer an audiophile myth. Dispel the myth if you want the answer you seek.

Quote
JA claimed signal-correlated activity in those last 8 LSBs.

I'd like to see a recording that shows this.



All of them that are well-made digital recordings.

That there is signal-related activity in  the low order bits is a truism.

Consider a recording of 20 Hz FS. The low order bits are flailing away to ensure that the signal has low noise and distortion.  Of course their flailing is signal-related. Reduce the signal to zero and their flailing will just be due to dither. Increase the signal and their flailing will be as required to preserve the low noise and distortion of the 20 Hz sine wave. Since there will now be quantization distortion that needs to be decorrelated, the flailing will change, subtly.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #360 –
Quote from: 2Bdecided on
...the correct answer is no. While it's true that you can twiddle the spectrum analysis to pull out this or that, there's an uncertainty-principle time/frequency relationship beyond which you cannot pull something out of noise. Beyond this, the thing isn't there. Or if it is, there's no way of determining whether it really is or isn't, so basically it isn't!


From this point on the debate should run pretty analogous to interpretations of quantum theory.

I'd say it is there and it is not there. Both is true. You can't pull it out (not there), but the actual noise would still be different if it hadn't been there. You cannot mathematically determine a point, where all correlation would have vanished, it never does completely. Capture a dirty old shellac disc with 24 bit resolution, with what you consider 30 dB of analog SNR and you still loose the fraction of a bit of signal, when you decimate that to 16 bit. We don't care, for a reason, but the question is better answered on the grounds of psycho-acoustics or practicability than math.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #361 –
Quote from: googlebot on
From this point on the debate should run pretty analogous to interpretations of quantum theory.
The uncertainty principle is from quantum theory.

I'm not entirely sure I should be drawing an analogy here, but it's close.


Quote
I'd say it is there and it is not there. Both is true. You can't pull it out (not there), but the actual noise would still be different if it hadn't been there. You cannot mathematically determine a point, where all correlation would have vanished, it never does completely. Capture a dirty old shellac disc with 24 bit resolution, with what you consider 30 dB of analog SNR and you still loose the fraction of a bit of signal, when you decimate that to 16 bit. We don't care, for a reason, but the question is better answered on the grounds of psycho-acoustics or practicability than math.
The re-quantisation measurably increases the noise floor. It's a small fraction of a dB when the original noise floor was already so (relatively) high, but it's there.

I think it reduces to this: was there (or could there have been) some detectable bit of signal sitting just above that original noise floor, that has now been swamped by the increase in the noise floor?

That depends on how you try to find the signal.

And whether you separately care about the noise floor.

As you say, psychoacoustics must come into a sensible discussion - but that's not the point I'm trying to track. In some signals, you realistically risk hiding some signal elements in a way that's trivial to demonstrate, whereas in others, you'd need a steady-state test signal and an hour long FFT to discover any difference in signal detection pre vs post quantisation. If we pick a sensible limit for the temporal integration time, we can calculate what amplitude of steady-state signal can be pulled out of the noise over that period of time. There's a trade-off between the two, and it does start to sound very Heisenberg-ish.

Brain very tired now!

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #362 –
Quote from: Stereoeditor on
Quote from: Ed Seedhouse on
But since it is a lossy codecs then by _definition_ "something" is removed.  Otherwise it wouldn't be a "lossy codecs".


Of course, and I expect that point is obvious to all on HA. But to the public at large, who have been told that even low-bit-rate satellite radio is "CD quality," the difference between lossless and lossy bit-rate reduction is not as clear as you might expect. To judge by some of the emails I receive, that difference isn't even as clear as I would expect among Stereophile's readership. That was the point I was making.


If what is removed cannot be discerned by the listener then so far as the listener is concerned the result is still of "CD" quality and it is perfetly reasonable to call it that.  The difference may be real, but if it is inaudible then it is irrelevant.  "A difference that makes no difference IS no difference".

Ed Seedhouse
VA7SDH

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #363 –
Quote from: 2Bdecided on
I think it reduces to this:



Assume a payload signal S1 and its innate noise floor, or uncertainty, N1.

When S1+N1 is properly quantised at a nominal resolution of n bits, this process
introduces a new uncertainty N2(n).

What is the minimal value of n that ensures that N1 > N2, '>' according to a
specific criterion?

The criterion follows from the payload signal
and the method of observation, e.g. is this about humans listening to
music, or is it about correlators digging for an a priori known signal.
This is not important as long as a criterion can be defined.

This is what TBD is trying to establish, working back from a given
recording with n=24.



No MSBs, DACs, monotonicity, or myths need to be mentioned or harmed in this discussion.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #364 –
I think that puts what I was trying to say in response to googlebot very succinctly.



It would be quite true to argue there are far more relevant/practical ways to look at this (I think Arny keeps trying to take us to correctly dithered signals and their noise floor - which is eminently sensible) - but I'm interested in those LSBs, what's in there, and what happens when they're chopped off (no dither).

I'm interested in the idea that sometimes some of them carry real information, while other times none of them carry any real information at all. I'm interested because this looks like a nice simple idea, while in fact it's rather complicated and difficult to pin down.

It's tenuously relevant if you're going to use noise-shaping and dither to convert to 16 bits, but it's highly relevant in the context of something like lossyWAV, or in considering the ill defined concept of "self dithering".


Anyway, here's a far simpler question: has anyone got a real-world 24-bit recording which has audibly signal-correlated information in the 8 LSBs?

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #366 –
Quote from: 2Bdecided on
Anyway, here's a far simpler question: has anyone got a real-world 24-bit recording which has audibly signal-correlated information in the 8 LSBs?

This is going to be tough. We know that any real recording has a noise floor higher than -96 dB so your 8 LSBs are going to have full-scale noise. Best case, there will also be full-scale distorted (amplitude aliased) signal. This gives us some hope. You can get a rough idea of how much hope by listening to the middle 8 bits. Your report on that is that sometimes you can make something out. So the relevant assessment is, do you think you could still make something out if you mixed that with white noise at the same (full-scale) level?

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #367 –
Quote from: Notat on
Quote from: 2Bdecided on
Anyway, here's a far simpler question: has anyone got a real-world 24-bit recording which has audibly signal-correlated information in the 8 LSBs?

This is going to be tough. We know that any real recording has a noise floor higher than -96 dB so your 8 LSBs are going to have full-scale noise. Best case, there will also be full-scale distorted (amplitude aliased) signal. This gives us some hope. You can get a rough idea of how much hope by listening to the middle 8 bits. Your report on that is that sometimes you can make something out. So the relevant assessment is, do you think you could still make something out if you mixed that with white noise at the same (full-scale) level?


What are the current constraints on "real world"?

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #368 –
Quote from: Notat on
So the relevant assessment is, do you think you could still make something out if you mixed that with white noise at the same (full-scale) level?
I think for a fair comparison you'd have to mix it before extracting those bits. Then it would act to dither the truncation. So you shouldn't have any audible difference, assuming it dithers it well enough.

I tried this...
a) 16-bit recording
b) truncate a to 8-bits
c) calculate a-b which gives the 8 LSBs of a.

The result is still often white noise, but it's easy enough to find real recordings where it isn't - i.e. where you can hear something related to the original signal.

Using such a recording, I carried out the above steps, and also did
d) add -40dB of white noise to a (i.e. somewhat above the level of the bits that will be removed)
e) truncate d to 8-bits
f) calculate d-e which gives the 8 LSBs of d

c has audible remnants of the original signal, f has nothing but white noise.

Which is what you'd expect - it's basic dither theory. If you don't add enough noise (e.g. -48dB peak-to-peak isn't enough), then some signal correlated elements can be heard in f.

To state the bleeding obvious, there's a huge audible difference between a and b, a and d, and a smaller but still obvious difference between d and e. In all cases the noise level is raised.

So, chopping off the LSBs caused an audible difference both in the case where those LSBs contained audible signal-correlated content, and in the case where they contained only white noise. Which again, is as expected, because you need about 5-bits to code white noise without changing the "sound" of it (i.e. to keep the quantisation noise below the noise "floor"), and in this case, it was only being given 1 or 2.

FWIW lossyWAV gets it right, and chops off bits (when possible), always without creating an audible difference.


Coming back to the beginning...
Quote from: Notat on
This is going to be tough. We know that any real recording has a noise floor higher than -96 dB so your 8 LSBs are going to have full-scale noise.
I'm not convinced that's always true. I know it's tough to get low real world noise floors across the entire audible band, but it is less tough to get fairly low noise levels at mid-frequencies. Either way, do this while recording some event that's briefly extremely loud, using state of the art equipment, and you might be there. I don't know - a cannon in an anechoic chamber?  Realistically, a loud instrument, recorded in a room with 20dB SPL noise or less? That'll get you 100dB+ SNR. You can easily cheat by putting the microphone very close to the instrument, and having only one instrument.

It doesn't sound like many real world recordings to me, but it might have turned up somewhere.

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #369 –
Quote from: Arnold B. Krueger on
What are the current constraints on "real world"?

Assume you are recording a source with non-lethal SPL (<120 dB SPL signal) in a real room (> 20 dB SPL noise floor).

(what 2BD said)

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #370 –
The 5-bit rule of thumb is new to me. Do you have a reference or can you elaborate?

Based on all of this, it seems like it might be possible to make a case that 24-bit resolution allows a recording to accurately reproduce the character of the noise floor of the equipment/environment in which the recording was made. Your experiments generally support the idea that it is unlikely there is any usable/audible information in those lower 8 bits related to the signal being recorded.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #371 –
In this 24bit discussion linked away from Slimdevices to "computeraudiophile" someone talks about hard evidence samples and by accident sells these!
http://www.computeraudiophile.com/content/...y#comment-72930
http://www.soundkeeperrecordings.com/format.htm

Anyone examined that bingo-bongo sounds?
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #372 –
Quote from: Wombat on
In this 24bit discussion linked away from Slimdevices to "computeraudiophile" someone talks about hard evidence samples and by accident sells these!
http://www.computeraudiophile.com/content/...y#comment-72930
http://www.soundkeeperrecordings.com/format.htm

Anyone examined that bingo-bongo sounds?
Do you mean "Kote Moun Yo?" from Equinox? Yes, just white noise in the last 8 bits.

"Dragon Boats" from Lift is far more interesting. Very strange stuff in the last 8 bits. Highly signal correlated. Very little noise. "But that's exactly what you're looking for David". Yeah, it would be if I was sure it was real. Problem is, there are strings of samples (12 here, 50 there, etc etc) scattered throughout this 24-bit file that are already quantised to 16-bits (not to mention the entirely 16-bit fades at either end). It's like the LSBs in this "24-bit" ADC were "sticky" - because, trust me, the signal from a microphone doesn't neatly quantise itself to 16-bits for 0.5ms - so something strange is happening here.

I do have the real "high resolution" disc purchased from that website. I keep meaning to check it (it might just be the upload that's "strange"), but I packed it away to move house nearly a year ago - and I'm still waiting to move!

EDIT: Here's the distribution of bit usage for the 8 LSBs of sr001-01-2496.wav (the downloaded extract of "Dragon Boats" from Lift, above)
[attachment=6397:sr001_01...t15_dist.gif]
I even removed the first second and last five seconds so as not to include the parts that have obviously been processed for upload (i.e. the fade out), though it made barely any difference to the graph because that's such a comparatively small section of the total length.

What you have in that graph is a count of how many times each possible value was found in the file. See the problem? Apart from during virtual silence (there's none in this extract) there's no reason why any specific value should appear more often than any other. But here, it's the values near 00000000 and 11111111 (255 in decimal) that occur far far more often than those in the middle (~128 decimal). A 24-bit value with 00000000 in the LSBs is really a 16-bit value. What this graph shows is that 24-bit values which are near 16-bit values are far more likely to occur that those which are midway between exact 16-bit values.

"Proper" 24-bit audio isn't like this. Something is wrong.


It's like sending someone out with a measuring tape to measure sticks, asking them to measure them to the nearest cm, but getting these results...
10
40
29
71
80
1
0
39
11
99

...you've got to wonder why the last digit is always 0, 1, or 9 - couldn't the person read numbers ending in 2-8?!

AFAICT the same thing is happening here with this supposedly 24-bit audio: The ADC seems to struggle measuring values that aren't near to 16-bit ones.

I have no idea how this happened. A best guess would be it's not the ADC at all, but some subsequent processing stage that's gone wrong. However, the recording engineer claims not to do any processing, even normalisation, because it degrades the sound.


EDIT2:
Here's the same graph for sr002-01-24192.wav ("Kote Moun Yo?" from Equinox, above)...
[attachment=6398:sr002_01_24192dist.gif]
There's nothing wrong here - all possible 8LSB values are used approximately the same number of times, exactly as you'd expect.

Cheers,
David.

P.S. I forgot to say: if you check the distribution of sample values for the full 24-bits of sr001-01-2496.wav, you get the typical double-sided poison-like distribution, but with _huge_ spikes at and around the 16-bit values. Here's an overall view...
[attachment=6399:sr001_01...ist_full.gif]

...and here's a zoom in showing the spikes clearly...
[attachment=6400:sr001_01...ist_zoom.gif]

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #373 –
Quote from: Notat on
The 5-bit rule of thumb is new to me. Do you have a reference or can you elaborate?
It's something I heard Bob Stuart say once. It seems to be about right, and it makes some sense from a typical simple minimum audible difference of ~ 1dB. I reckon it's closer to 0.5dB and depends on how you measure the noise, but it's something like that.

Sorry, no reference.

Cheers,
David.

"Audiophile" listening event @ Definitive Audio in Seattle

Reply #374 –
Quote from: 2Bdecided on
"Dragon Boats" from Lift is far more interesting. Very strange stuff in the last 8 bits. Highly signal correlated. Very little noise. "But that's exactly what you're looking for David". Yeah, it would be if I was sure it was real. Problem is, there are strings of samples (12 here, 50 there, etc etc) scattered throughout this 24-bit file that are already quantised to 16-bits (not to mention the entirely 16-bit fades at either end). It's like the LSBs in this "24-bit" ADC were "sticky" - because, trust me, the signal from a microphone doesn't neatly quantise itself to 16-bits for 0.5ms - so something strange is happening here.


I just tried that piece of music and me as noob also wonders that it fades in and out very quick so you don´t have a chance to hear recorded slience. Since my player only plays 96kHz natively i tried 2 soxed versions with 44.1 and 96kHz on my main system. I was even trying listening as loud as my system can but only found the 44.1 version the same sounding.

Edit: I tried both samples offered on their sample page
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!