IPB

Welcome Guest ( Log In | Register )

16 Pages V  « < 12 13 14 15 16 >  
Reply to this topicStart new topic
"Audiophile" listening event @ Definitive Audio in Seattle, Atkinson to demonstrate "evils of MP3"
greynol
post Mar 3 2011, 21:26
Post #326





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



Has anyone attempted this with the samples provided which are actually on-topic to this discussion?


--------------------
I should publish a list of forum idiots.
Go to the top of the page
+Quote Post
googlebot
post Mar 3 2011, 21:31
Post #327





Group: Members
Posts: 698
Joined: 6-March 10
Member No.: 78779



I remember JA offering recordings with non-random content below 16 bit, but I do not remember seeing any actual links. Have I missed anything?
Go to the top of the page
+Quote Post
greynol
post Mar 3 2011, 21:34
Post #328





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



http://www.hydrogenaudio.org/forums/index....st&p=743399


--------------------
I should publish a list of forum idiots.
Go to the top of the page
+Quote Post
Stereoeditor
post Mar 3 2011, 21:55
Post #329





Group: Members
Posts: 227
Joined: 24-April 09
Member No.: 69239



QUOTE (googlebot @ Mar 3 2011, 14:38) *
Cutting LSBs is fine, but cutting MSBs leads to serious distortion due to clipping.


Not in my experience. Have you tried this for yourself with a recording that has correlated information in bits 17-24? I think the word "clipping" is being misused here.

As "Notat" wrote in a recent message: "If we hear anything that is signal dependent in the LS bits, that's very strong evidence that there's information in those bits - reduced entropy indicates presence of information."

That's the point I was making at the AES workshop: that with _some_ 24-bit recordings, there are valid signal data in bits 17-24.

John Atkinson
Editor, Stereophile

This post has been edited by Stereoeditor: Mar 3 2011, 22:00
Go to the top of the page
+Quote Post
googlebot
post Mar 3 2011, 22:35
Post #330





Group: Members
Posts: 698
Joined: 6-March 10
Member No.: 78779



QUOTE (Stereoeditor @ Mar 3 2011, 21:55) *
I think the word "clipping" is being misused here.


The phenomenon is not only something comparable to clipping but exactly the correct term. It might be easier to comprehend for you when you realize that stripping the MSBs is equivalent to digital amplification (multiplication by 2^n, n=number of bits), hard into 0dB, followed by attenuation (division by 2^n) back to the original level. This is a paramount example for digital clipping, there is not the slightest misuse of the word.

This post has been edited by googlebot: Mar 3 2011, 23:24
Go to the top of the page
+Quote Post
Arnold B. Kruege...
post Mar 3 2011, 23:27
Post #331





Group: Members
Posts: 3700
Joined: 29-October 08
From: USA, 48236
Member No.: 61311



QUOTE (Stereoeditor @ Mar 3 2011, 15:55) *
QUOTE (googlebot @ Mar 3 2011, 14:38) *
Cutting LSBs is fine, but cutting MSBs leads to serious distortion due to clipping.


Not in my experience.


You need to open your eyes up and look at what happens to *any* waveform when you start stripping off the MSBs.

QUOTE
Have you tried this for yourself with a recording that has correlated information in bits 17-24?


Yup.

QUOTE
I think the word "clipping" is being misused here.


I think that if you actually look at a wave that has had the first few MSBs zeroed out, I'll bet that you'll say to yourself: "Why that wave has been clipped"!
Go to the top of the page
+Quote Post
Stereoeditor
post Mar 4 2011, 01:45
Post #332





Group: Members
Posts: 227
Joined: 24-April 09
Member No.: 69239



QUOTE (googlebot @ Mar 3 2011, 16:35) *
QUOTE (Stereoeditor @ Mar 3 2011, 21:55) *
I think the word "clipping" is being misused here.


The phenomenon is not only something comparable to clipping but exactly the correct term. It might be easier to comprehend for you when you realize that stripping the MSBs is equivalent to digital amplification (multiplication by 2^n, n=number of bits), hard into 0dB, followed by attenuation (division by 2^n) back to the original level. This is a paramount example for digital clipping, there is not the slightest misuse of the word.


Thank you. Yes, it does appear I was incorrect. The question then becomes: is the residue random in nature, ie white noise, which when clipped is unchanged, or is it still signal-related, in which case the 8 LSBs did contain audio data and not random bit switching?

John Atkinson
Editor, Stereophile
Go to the top of the page
+Quote Post
Notat
post Mar 4 2011, 06:32
Post #333





Group: Members
Posts: 581
Joined: 17-August 09
Member No.: 72373



QUOTE (googlebot @ Mar 3 2011, 14:35) *
The phenomenon is not only something comparable to clipping but exactly the correct term. It might be easier to comprehend for you when you realize that stripping the MSBs is equivalent to digital amplification (multiplication by 2^n, n=number of bits), hard into 0dB, followed by attenuation (division by 2^n) back to the original level. This is a paramount example for digital clipping, there is not the slightest misuse of the word.

Not exactly correct. Clipping should imply saturation for samples outside the allowed range. What happens here is that out-of-range samples are wrapped back into range. Sort of an amplitude aliasing thing. Sounds very nasty but does not totally whiten or clobber the information.
Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 10:24
Post #334


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



Notat is correct. It's not clipping at all.

If you clipped it like this, you'd certainly hear something related to the music afterwards, whatever the original contents of the 8LSBs.

Cheers,
David.
Go to the top of the page
+Quote Post
googlebot
post Mar 4 2011, 11:07
Post #335





Group: Members
Posts: 698
Joined: 6-March 10
Member No.: 78779



Could someone enlighten me?

1. If I amplify a sequence of samples n bits over what their containers can hold, I get digital clipping. The top n bits fall over board. I guess no one wants to challenge that.
2. If I attenuate the same sequence afterwards by the same amount, the resulting sequence will be identical to the original sequence with (n) MSBs blanked and (total - n) LSBs untouched.

How does 2. invalidate the term "clipping"?

If the DAC has a little headroom above its rated bitdepth (many do) 1. and 2. should lead to comparable spectral components after DA conversion.

This post has been edited by googlebot: Mar 4 2011, 12:03
Go to the top of the page
+Quote Post
Arnold B. Kruege...
post Mar 4 2011, 11:51
Post #336





Group: Members
Posts: 3700
Joined: 29-October 08
From: USA, 48236
Member No.: 61311



QUOTE (Notat @ Mar 4 2011, 00:32) *
QUOTE (googlebot @ Mar 3 2011, 14:35) *
The phenomenon is not only something comparable to clipping but exactly the correct term. It might be easier to comprehend for you when you realize that stripping the MSBs is equivalent to digital amplification (multiplication by 2^n, n=number of bits), hard into 0dB, followed by attenuation (division by 2^n) back to the original level. This is a paramount example for digital clipping, there is not the slightest misuse of the word.

Not exactly correct. Clipping should imply saturation for samples outside the allowed range. What happens here is that out-of-range samples are wrapped back into range. Sort of an amplitude aliasing thing. Sounds very nasty but does not totally whiten or clobber the information.


I see your point, and its important as far as it goes. However, we're still talking about massive ruination of the wave form and applying huge amounts of nonlinear distoriton. I know of no standard that accepts or recommends applying massive nonlinear distortion to a signal in order to recover information about small components of it.

BTW there are natural situations that have similar results such as old-style ladder DACs that are missing a lot of codes that are next to each other. I don't recall anybody ever recomending using broken DACs to measure low level detail like crossover distortion. The usual recommendation was to fix them so that they were linear and monotonic!
Go to the top of the page
+Quote Post
spoon
post Mar 4 2011, 12:01
Post #337


dBpowerAMP developer


Group: Developer (Donating)
Posts: 2745
Joined: 24-March 02
Member No.: 1615



Removing MSB bits will result in a very distorted sound wave with large transients (square waves), a simple representation is a 8 bit signal (0-255 for simplicity with no -), if the upper 8th bit is removed then anything above 128 will jump down to signal-128, so 126 and 127 would be ok, 128 would become 0 and 129 would become 1.

This post has been edited by spoon: Mar 4 2011, 12:04


--------------------
Spoon http://www.dbpoweramp.com
Go to the top of the page
+Quote Post
WernerO
post Mar 4 2011, 12:04
Post #338





Group: Members
Posts: 73
Joined: 21-November 06
Member No.: 37858



QUOTE
Notat is correct. It's not clipping at all.


It is equivalent to listening to the isolated quantisation distortion introduced by truncating the 24b source
material to 16b.

If the residue is / sounds like just noise this means that the 24b original had a sub-16b innate noise floor, overwhelming any payload signal in the lower 8 bits. And that seems to be exactly what you get with the sample referred to above, admittedly tried only in a noisy office environment and with cheap open headphones.

(edit)

Doing the same with 24b versus 8b reveals a crackling residue. Re-doing 24b versus 16b now (lunchtime, hence quieter environment and louder replay levels) reveals a high-pitched tinnitus-like signal that is keyed-on and -off, mostly in the right channel. So the residue is not strictly white.



This post has been edited by WernerO: Mar 4 2011, 12:11
Go to the top of the page
+Quote Post
Dirk95100
post Mar 4 2011, 12:15
Post #339





Group: Members
Posts: 33
Joined: 15-October 10
Member No.: 84639



I was wondering, how am I able to set the top msb to 0?
Is there a tool that can do that?
I mean without amplifing the signal over 0dB and then reducing gain.
Go to the top of the page
+Quote Post
WernerO
post Mar 4 2011, 12:18
Post #340





Group: Members
Posts: 73
Joined: 21-November 06
Member No.: 37858



Truncate to 16 bit. Expand again to 24 bit (i.e. all 8 LSBs are now zero). Subtract from 24 bit source (all 16 MSBs are now zero).
Go to the top of the page
+Quote Post
Dirk95100
post Mar 4 2011, 13:06
Post #341





Group: Members
Posts: 33
Joined: 15-October 10
Member No.: 84639



QUOTE (WernerO @ Mar 4 2011, 12:18) *
Truncate to 16 bit. Expand again to 24 bit (i.e. all 8 LSBs are now zero). Subtract from 24 bit source (all 16 MSBs are now zero).

Thanks Werner.
I tried it with some drum loops I got from the net and I hear besides hudge amounts of noise some drum sounds to.
Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 14:07
Post #342


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



QUOTE (greynol @ Mar 3 2011, 20:26) *
Has anyone attempted this with the samples provided which are actually on-topic to this discussion?
QUOTE (greynol @ Mar 3 2011, 20:34) *

Yes.

Bits 17-24 sound like white noise to me.

I've uploaded them next to the original...
http://www.hydrogenaudio.org/forums/index....showtopic=86738

(It's still an 88.2kHz 24-bit file, though there are a lot of zeros in there, hence the comparatively small FLAC filesize for a noise-like signal which hits digital full scale - wasted_bits is very useful!).

Cheers,
David.

EDIT: I posted this without seeing the last 7 posts in this thread.

This post has been edited by 2Bdecided: Mar 4 2011, 14:11
Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 14:29
Post #343


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



QUOTE (WernerO @ Mar 4 2011, 11:04) *
It is equivalent to listening to the isolated quantisation distortion introduced by truncating the 24b source material to 16b.
I agree.

QUOTE
Doing the same with 24b versus 8b reveals a crackling residue. Re-doing 24b versus 16b now (lunchtime, hence quieter environment and louder replay levels) reveals a high-pitched tinnitus-like signal that is keyed-on and -off, mostly in the right channel. So the residue is not strictly white.
I don't think your software has done what you think it has. See the sample I posted.

FWIW In Cool Edit Pro, if you convert 24>16, and then subtract the 16 from the 24, you don't just get the 8LSBs. The undithered 24>16 in Cool Edit Pro isn't quite the simple bit discarding that you might expect, but actually rounds all values up (except zero!).

Even so, you still get something which sounds like it should (except for the jumps in amplitude on any zeros - which is still inaudible in a noisy signal), even though mathematically it's wrong.

Cheers,
David.
Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 14:34
Post #344


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



QUOTE (Dirk95100 @ Mar 4 2011, 12:06) *
I tried it with some drum loops I got from the net and I hear besides hudge amounts of noise some drum sounds to.
I don't think there's any doubt that you'll hear something for artificial signals, generated at 24-bits.

You can generate an artificial signal with real signal-correlated data to as many bits as you want.

A sine wave generated accurately to 32-bits or 64-bits would still have something signal-correlated in the last 8 bits, showing that 24 or even 56 bits just aren't enough. wink.gif

Cheers,
David.
Go to the top of the page
+Quote Post
WernerO
post Mar 4 2011, 15:14
Post #345





Group: Members
Posts: 73
Joined: 21-November 06
Member No.: 37858



Bugger. My tool's manual claims it truncates, but the resultant difference file sounds a bit dirtier than yours, which indeed is a rather clean kind of noise. To Be Characterised.
Go to the top of the page
+Quote Post
Dirk95100
post Mar 4 2011, 15:30
Post #346





Group: Members
Posts: 33
Joined: 15-October 10
Member No.: 84639



QUOTE (2Bdecided @ Mar 4 2011, 14:34) *
QUOTE (Dirk95100 @ Mar 4 2011, 12:06) *
I tried it with some drum loops I got from the net and I hear besides hudge amounts of noise some drum sounds to.
I don't think there's any doubt that you'll hear something for artificial signals, generated at 24-bits.

You can generate an artificial signal with real signal-correlated data to as many bits as you want.

A sine wave generated accurately to 32-bits or 64-bits would still have something signal-correlated in the last 8 bits, showing that 24 or even 56 bits just aren't enough. wink.gif

Cheers,
David.


It was a recording of the famous Amen drum loop, so its not generated but recorded from LP.
Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 15:49
Post #347


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



QUOTE (Dirk95100 @ Mar 4 2011, 14:30) *
It was a recording of the famous Amen drum loop, so its not generated but recorded from LP.
I never knew that drum track was called that! The wikipedia page is a goldmine!

I'm surprised there's anything important in the 12th bit, never mind the 20th. I suspect something strange is happening, but who knows.

Cheers,
David.
Go to the top of the page
+Quote Post
Notat
post Mar 4 2011, 16:44
Post #348





Group: Members
Posts: 581
Joined: 17-August 09
Member No.: 72373



QUOTE (WernerO @ Mar 4 2011, 04:04) *
It is equivalent to listening to the isolated quantisation distortion introduced by truncating the 24b source
material to 16b.

If the residue is / sounds like just noise this means that the 24b original had a sub-16b innate noise floor, overwhelming any payload signal in the lower 8 bits.

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. Just because we don't hear anything doesn't mean there's nothing important there. This is a non-linear process and so difficult to model how the signal is transformed. There are processes which will make information sound like white noise. Encryption is the obvious example.

Also be careful when discussing noise floor. There almost always is useful information below the noise floor.

This post has been edited by Notat: Mar 4 2011, 16:45
Go to the top of the page
+Quote Post
WernerO
post Mar 4 2011, 17:29
Post #349





Group: Members
Posts: 73
Joined: 21-November 06
Member No.: 37858



QUOTE (Notat @ Mar 4 2011, 16:44) *
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.

Go to the top of the page
+Quote Post
2Bdecided
post Mar 4 2011, 17:48
Post #350


ReplayGain developer


Group: Developer
Posts: 5104
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



QUOTE (WernerO @ Mar 4 2011, 16:29) *
QUOTE (Notat @ Mar 4 2011, 16:44) *
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.

Agreed.

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 wink.gif.

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.
Go to the top of the page
+Quote Post

16 Pages V  « < 12 13 14 15 16 >
Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 28th August 2014 - 08:36