24/88.2 FLAC clip from Handel's Messiah, Linn Records Options

[mixminus1]

From this thread

 13._.Chorus.For.unto.us.a.child.is.born.1min21sec_1min45sec.flac ( 7.74MB ) Number of downloads: 863

[2Bdecided]

Here are bits 17-24 (only) from each sample of the above file, moved to bits 2-9 so you can hear them.

Warning: quite loud! (Far louder than the original file).
Warning: DC offset (bit 1 not used, hence all values are positive).

 13._.Chorus.For.unto.us.a.child.is.born.1min21sec_1min45sec_proc.flac ( 4.52MB ) Number of downloads: 129


EDIT: generated in MATLAB using manual bit shifting. DSP or assembler would have been far more efficient, but I believe the MATLAB result is correct.

This post has been edited by 2Bdecided: Mar 4 2011, 14:30

[bandpass]

Spectrogram of above, showing nothing but white noise (and the DC offset).



-bandpass

[Nick.C]

LWCDF file attached [File removed - obsolete] (i.e. numerical difference between lossless and lossyWAV processed audio) - processed with adaptive shaping OFF using a version of lossyWAV which will always keep 16 bits, therefore exactly 8 bits removed. File is ReplayGained (+64.81dB)

CODE

lossyWAV beta 1.2.4b, Copyright (C) 2007-2011 Nick Currie. Copyleft.
Processed : 05/03/2011 20:14:44
Settings  : --quality extraportable --adaptive off
Filename  : 13._.Chorus.For.unto.us.a.child.is.born.1min21sec_1min45sec.wav
File Info : 88.20kHz; 2 channel; 24 bit, 00:24.87, 12.55MiB
Results   : 8.0000 bits; 21.81x; 00:01.14; [F]

Sample least significant bit (lsb) distribution.
+---+------------------------------------+------------------------------------+
|Bit|Input sample lsb distribution       |Output sample lsb distribution      |
+---+-------+----------------------------+-------+----------------------------+
|NUL|  0.00%|............................|  0.03%|............................|
|  0| 49.97%|OOOOOOOOOOOOOO..............|  0.00%|............................|
|  1| 25.02%|OOOOOOO.....................|  0.00%|............................|
|  2| 12.50%|OOOO........................|  0.00%|............................|
|  3|  6.25%|OO..........................|  0.00%|............................|
|  4|  3.13%|O...........................|  0.00%|............................|
|  5|  1.56%|............................|  0.00%|............................|
|  6|  0.79%|............................|  0.00%|............................|
|  7|  0.39%|............................|  0.00%|............................|
|  8|  0.20%|............................| 49.81%|OOOOOOOOOOOOOO..............|
|  9|  0.10%|............................| 25.10%|OOOOOOO.....................|
| 10|  0.05%|............................| 12.55%|OOOO........................|
| 11|  0.02%|............................|  6.27%|OO..........................|
| 12|  0.01%|............................|  3.13%|O...........................|
| 13|  0.01%|............................|  1.57%|............................|
| 14|  0.00%|............................|  0.78%|............................|
| 15|  0.00%|............................|  0.40%|............................|
| 16|  0.00%|............................|  0.20%|............................|
| 17|  0.00%|............................|  0.10%|............................|
| 18|  0.00%|............................|  0.05%|............................|
| 19|  0.00%|............................|  0.02%|............................|
| 20|  0.00%|............................|  0.00%|............................|
| 21|  0.00%|............................|  0.00%|............................|
| 22|  0.00%|............................|  0.00%|............................|
| 23|  0.00%|............................|  0.00%|............................|
+---+-------+----------------------------+-------+----------------------------+

This post has been edited by Nick.C: Apr 24 2011, 20:20

[bandpass]

generated in MATLAB using manual bit shifting. DSP or assembler would have been far more efficient, but I believe the MATLAB result is correct.

Here's how to do it with SoX (by pretending that the 2 24-bit channels are 6 8-bit channels):

CODE

sox input.flac -twav -|sox -eun -b8 -c6 - output.flac remix 1 4

or for just a spectrogram:

CODE

sox input.flac -twav -|sox -eun -b8 -c6 - -n remix 1 4 spectrogram -z 50

[googlebot]

Clever idea!

[IgorC]

Here are bits 17-24 (only) from each sample of the above file, moved to bits 2-9 so you can hear them.

Warning: quite loud! (Far louder than the original file).
Warning: DC offset (bit 1 not used, hence all values are positive).

 13._.Chorus.For.unto.us.a.child.is.born.1min21sec_1min45sec_proc.flac ( 4.52MB ) Number of downloads: 129


EDIT: generated in MATLAB using manual bit shifting. DSP or assembler would have been far more efficient, but I believe the MATLAB result is correct.

Don't want to be harsh but what purpose of such extraction? Point me if I am wrong but separated less significant bits (17-24) don't mimic the original signal and contain pure quantization differences/errors (noise).

P.S. I don't know if it's placebo but I hear some background intelligible sounds

This post has been edited by IgorC: Mar 6 2011, 13:26

[bandpass]

Don't want to be harsh but what purpose of such extraction? Point me if I am wrong but separated less significant bits (17-24) don't mimic the original signal and contain pure quantization differences/errors (noise).

P.S. I don't know if it's placebo but I hear some background intelligible sounds

Whatever should be in those LSBs, it clearly shouldn't be white noise.

P.S. If you listen to white noise for long enough, you'll eventually hear the sound of monkeys typing out the works of Shakespeare...

[IgorC]

Whatever should be in those LSBs, it clearly shouldn't be white noise.

Not white, but quantization noise.

*Stupid* Example:

Imagine you've done 3 measurements with 24-bit tool of length measurement. 10.3/12.2/13.1 cm . As you can see it's increasing function
After that you've done exactly the same 3 measurements but now with 16 bit tool 10/12/13 cm. It's still the same function with the same characteristics but less precise.

Now extract those (17-24) bits >
10.3 -10 = 0.3
12.2 -12 = 0.2
13.1 - 13 = 0.1

Your final function is 0.3/0.2/0.1 - totally different function with decreasing character and has nothing in common with original function!!!



So, I'm not saying that LSBs (17-24) are useless and don't contain useful data but I think (and might be wrong) that _without_ MSBs those bits have nothing in common with original signal.

This post has been edited by IgorC: Mar 6 2011, 14:41

[bandpass]

So, I'm not saying that LSBs (17-24) are useless and don't contain useful data but I think (and might be wrong) that _without_ MSBs those bits have nothing in common with original signal.

Take this example: a 1500Hz sine wave is rendered at 48k in bits 1-8 of a 16 bit signal; bits 9-16 are set to white noise. A spectrogram of bits 9-16 obviously has no features at this point. The 16-bit signal is then multiplied linearly by 0.5. There is now some information relating to the sine wave in bit 9 and the spectrogram of bits 9-16 looks like this:



i.e. not white noise and showing the presence of frequencies that do relate to the original function (they are multiples thereof).

[2Bdecided]

Don't want to be harsh but what purpose of such extraction?

In this thread...
http://www.hydrogenaudio.org/forums/index....showtopic=86649
...JA claimed to find "signal-correlated activity in the 8 LSBs" of some recordings...
http://www.hydrogenaudio.org/forums/index....mp;#entry746446

I was giving others the opportunity to listen to the 8 LSBs of this recording...
http://www.hydrogenaudio.org/forums/index....st&p=746801

Cheers,
David.

[2Bdecided]

Whatever should be in those LSBs, it clearly shouldn't be white noise.

?

When you quantise a signal where the entire signal spectrum is well above the quantisation level, the difference between the original and quantised version (i.e. the quantisation noise) is white noise.

It obviously has something to do with the original signal (it's the last 8 bits of it in this case), but whether it is any way correlated or even meaningfully related to the original signal is another question.

Anyway, I think the discussion would be better continued in the original thread...
http://www.hydrogenaudio.org/forums/index....showtopic=86649

Cheers,
David.

[IgorC]

2Bd,
Thanks for pointing me.

Well, Bandpass has nice screenshot that apparently shows the correlation.

This post has been edited by IgorC: Mar 7 2011, 13:06

[2Bdecided]

Well, Bandpass has nice screenshot that apparently shows the correlation.

IMO synthetic files aren't really relevant. I can create a synthetic file that shows signal correlated information in the last 8-bits of a 64-bit file.

Cheers,
David.