Dithering to 24 bits for mastering?

[monomer]

Originally Posted by Lupo

What do you mean by signal theory and raw information? Theory vs real world implementation?

No, its different.
I see the process as a direct manipulation of bits which alows me to step outside the notion of signal theory altogether.
Since dither is based in signal theory i apparently wasn't even talking about dither!
I was talking about single sample bit manipulations.
But these type of manipulations show that there is information present which is not considered by signal theory.

[kjg]

Originally Posted by monomer

Try this test:

Take your 24 bit sines file.
Make 2 copies of it.
I'll call them A, B and C.
Now dither B to 8 bits and truncate C to 8 bits
There is a difference between these files and i may have erroneously called this difference Dither.
But let's call it 'Extra information'.
I call it this way because after dithering the B file is truncated, just like C was.
Therefore any difference between these would have to have come from the dithering process.

No. The difference between these two is:
1. the dither noise (in B)
2. the truncation distortion (in C)

Because they are indeed exclusive. Properly dithered >> no truncation distortion correlated to the signal.

Originally Posted by Bob Olhsson

Dither makes truncation sound like noise.

You are nulling these two files (B & C) and think that the signal you see is a result of the dither. It is not. The noise is a result of the dithering (as seen in nulling B with original A), and the "extra information" is just truncation distortion, which is indeed correlated with the original signal. And that is exactly why you want to dither before truncating. To completely remove any correlation in the distortion that occurs when truncating. That is why you modulate that LSByte with noise.

Originally Posted by Bob Olhsson

Dither makes truncation sound like noise.

Originally Posted by monomer

Now mix an inverted B with original A.
You state that file A would now only contain noise (as per your test).

And that is indeed the case.

Copy file C into clipboard and mix it inverted with file B
This leaves only the 'extra information that was added by the dithering process.

No. This leaves the dither noise added to B plus the truncation distortion from C. No extra information. Dither noise and truncation distortion, because you mixed a properly dithered (and then truncated) file with an inverted undithered truncated file.

The truncation distortion (of C) is correlated with the original signal, and that is the trace of signal you erroneously call "extra information".



best,
Klaas-Jan

[kjg]

Originally Posted by monomer

But these type of manipulations show that there is information present which is not considered by signal theory.

No they don't. They show the truncation distortion of C plus the dither noise of B.

I wasn't expecting the party to last this long...

[Nordenstam]

It's been going on for a while now.. :D

Originally Posted by monomer

No, its different.
I see the process as a direct manipulation of bits which alows me to step outside the notion of signal theory altogether.
Since dither is based in signal theory i apparently wasn't even talking about dither!
I was talking about single sample bit manipulations.
But these type of manipulations show that there is information present which is not considered by signal theory.

Signal theory is about signals - anything possible to create. The tests done in sound forge is about real world implementations of those theories.

Let me try to say it in everyday mastering language. To master, we work at high bit depths with the adding of dither noise being the penultimate process prior to 16 bit truncation. That's it. Add dither, truncate and that's the last thing to be done to the data before burning master disk. To create an additional truncated 16 bit file and deliberate mix this with the 16 bit master is both unnecessary and aesthetically wrong in most peoples ears. Adding a truncated copy got nothing to do with the normal use of bit depth reduction.


This can also been seen when starting your procedure. When adding the truncation C file with the dither B file, you've got two options. If you mix with normal polarity on both, you're adding another copy of the sine waves on top of the already existing sines in the dithered file. You're doubling the sine waves signal strength. If you mix with reversed polarity on the truncation file, you'll cancel out the sine waves, they'll disappear. This is crucial. You're either removing the sines or adding another copy of them on top of the already existing sines. (in addition to the positive or negative truncation being added) If you open the frequency analyzer (alt-8) you'll clearly see the high level sine waves either disappear or increase in level.

It can be also be shown when cancelling the new mixed B file against the A file. With the extra layer of sine waves and truncation distortion added, mixing with the 24 bit file will raise the level of the sines (if you added positive truncation distortion while mixing the B and C files) or keep the level of the sine waves intact(if you added a negative of the truncation distortion while mixing the B and C files). If the B file you created by mixing with C was a true bit depth reduced copy of the 24 bit files, nulling against the 24 bit file should cancel the sine waves.

Edit: PS: if you start with a 0dB sine wave, it'll be much easier to see - even in waveform view. It'll either clip or disappear when you try to mix the B and C files.

[monomer]

Originally Posted by Lupo

It's been going on for a while now.. :D



Signal theory is about signals - anything possible to create. The tests done in sound forge is about real world implementations of those theories.

I understand what you mean and i concur that everything you propose is right.
But it is only right within the mathematical framework we call signal theory.

Signal theory provides in dealing with signals.
Signals are information changed over time.
Signals therefore imply a spectrum.
Information by itself does not imply this.
Information can ruthlessly deal with a single sample without concerning itself with it's effect on the signal.
Information deals with rows of ones and zeroes, anything goes.
What i am doing applies only to the information (so without the notion of time).

I'm currently looking into coding a vst plugin that will remove the noise from the 'residu noise' data to show that the old leftover lower bits are there unchanged.
I will do that by zeroing out only one bit in every sample.

[dcollins]

Originally Posted by kjg

I wasn't expecting the party to last this long...

I can't tell if the entropy is increasing, or not................


DC

[staudio]

How would one find out whether Logic is truncating to 24 bit when doing a mix down? It is not documented anywhere in the official reference materials.

[Bob Olhsson]

Dave's here...

Hi Dave!

[jayfrigo]

Originally Posted by dcollins

I can't tell if the entropy is increasing, or not................


DC

Perhaps James A. Yorke could help...

[okydoky]

Originally Posted by jayfrigo

Perhaps James A. Yorke could help...

hi,

he was prepared to provide an analysis. however, unfortunately, he died of boredom while reading posts 128 through 185.

accordingly, oky****'s 3rd theorum applies. it states as follows:

the degree of incoherence of the noise signal in a gearslutz thread increases in inverse proportion to the quality of the software used to support the various arguments.



right.

[dcollins]

Originally Posted by oky****

the degree of incoherence of the noise signal in a gearslutz thread increases in inverse proportion to the quality of the software used to support the various arguments.

It's a well-known relationship.

c = b log2 (1 + s)

Where c is the cost of the software, b is the brightness of the dither signal and and s is the sillyness of the argument.


DC

[Nordenstam]

Originally Posted by staudio

How would one find out whether Logic is truncating to 24 bit when doing a mix down? It is not documented anywhere in the official reference materials.

Have some audio with a peak value at at -150 to -190 dBFS. If the file is empty after mixdown, it truncates. If there's a noise, there's dither. Boosting the noise say 100dB should reveal that the audio is inside the noise floor.

[Nordenstam]

Originally Posted by monomer

I understand what you mean and i concur that everything you propose is right.
But it is only right within the mathematical framework we call signal theory.

Its' not only right within the theoretical framework. It's right when creating 16 bit master disks and that's the most imparting thing for most folks when it comes to dither. It's also right when treating any conceivable real world signal that is at least a few hundred samples long..

Originally Posted by monomer

Signal theory provides in dealing with signals.
Signals are information changed over time.
Signals therefore imply a spectrum.
Information by itself does not imply this.
Information can ruthlessly deal with a single sample without concerning itself with it's effect on the signal.

A single sample by itself is nothing. It's as theoretical as the "point" idea in math - it's a good thing to have among the tools, but it can not exist in the real world. As you noted, it does not have a spectrum per se - because there is no information there! The point could be anything, anywhere in the universe. It's impossible to tell what or where a point is without a reference. If you're going to measure the point or place the point somewhere, you need at least another point to compare to. If there are two sample points with different values, there will be some change in energy level. That energy change will have a spectrum, so we're into the "signal theory land". The spectrum of the change in level is the information contained therein.

In any case, you'll always listen to a sampled system that is continuos. I wager to say there's never less than at least a few hundred samples. That would be a few millisecs of samples at 44.1 sampling - probably the shortest snippet of sound it'll be possible to find in any sample bank ever made. Most often times, millions and billions of samples comes in a row. The very idea of digital audio presumes the samples to come as a train with the sample points evenly spaces.

Originally Posted by monomer

Information deals with rows of ones and zeroes, anything goes.
What i am doing applies only to the information (so without the notion of time).

If working on more than a single sample, the information is manipulated across a stretch of time.

Originally Posted by monomer

I'm currently looking into coding a vst plugin that will remove the noise from the 'residu noise' data to show that the old leftover lower bits are there unchanged.

You have to sell the VST plug for 10K dollars. That way, your arguments will weight in heavy!

Anyway; the regular use of sound forge with nulling etc can prove anything on this subject.


This have been dragging on way too far now. If it's possible to share knowledge, fine, but if that doesn't work - there's no point in going on for either of us. I'm off to the after party.


Cheers,

Andreas

[monomer]

Originally Posted by Lupo

This have been dragging on way too far now. If it's possible to share knowledge, fine, but if that doesn't work - there's no point in going on for either of us. I'm off to the after party.


Cheers,

Andreas

But, but,..,
But the twins are about to perform!


ceeya!

[UnderTow]

Originally Posted by kjg

so when you say LSB you mean Least Significant Bit? I meant Least Significant Byte.

I think you are both wrong.

From my understanding it isn't about a fixed number of bits. Rather the dither needs to cover all the bits being truncated from the source.

When, for instance, going from 24 bit to 8 bit, you need more than a Byte worth of dither.

Alistair

[kjg]

Originally Posted by UnderTow

I think you are both wrong.

From my understanding it isn't about a fixed number of bits. Rather the dither needs to cover all the bits being truncated from the source.

Of course. Thanks for clarifying.
It isn't about a fixed number of bits. That was just about the common situation of going from 24 to 16 bit.

When, for instance, going from 24 bit to 8 bit, you need more than a Byte worth of dither.

Alistair

Yes, in that case it would be the two LSBytes.
Or when truncating from 24 to 20 only the least significant nibble.

I think "nibble" is my favorite term in computer land, btw :D

[monomer]

Originally Posted by kjg

I think "nibble" is my favorite term in computer land, btw :D

agreed...
I sometimes prefer saying 'two nibbles' instead of 'byte'.

[kjg]

Originally Posted by monomer

agreed...
I sometimes prefer saying 'two nibbles' instead of 'byte'.

Yes. Especially when the twins are performing..

Maybe you want to ask a mod to rename the thread to "Dithering to 6 nibbles for mastering?".

[Nordenstam]

V-twins?

The dither covers all the bits that are to be removed in truncation, plus one to three bits of the remainding data. Ie, 9-12 bits of dither for 24->16 bit conversion, depending on shaping.

[Kees de Visser]

Originally Posted by kjg

so when you say LSB you mean Least Significant Bit? I meant Least Significant Byte.

That's interesting. According to Wikipedia you're right to distinguish between LSB and lsb.

In computing, the least significant bit (lsb) is the bit position in a binary integer giving the units value,
...
LSB, in all capitals, can also stand for "Least Significant Byte". The meaning is parallel to the above: it is the byte (or octet) in that position of a multi-byte number which has the least potential value.

However in literature about dither (and other digital audio topics) I've always seen LSB (in capitals) meaning least significant bit. E.g. "recommended dither is 2 LSB peak-to-peak triangular", which obviously isn't 2 LSBytes
IMO the use of Bytes make sense for file formats since most audio these days is somehow computer related, which wasn't the case e.g. with digital tape recorders like DASH.
I think it's a good idea to stick to the wikipedia definitions.

[UnderTow]

Originally Posted by Lupo

It's very strange that you've got the same odd result. The 8 bit converter does its job as evident on the FFT display. Copying from this copies the 8 bit data. Though I did save it to files now too, just to be sure.

Your "noise residual" file also has these weird "sineish looking" waveforms. As I, said it isn't audible as a sine and it isn't a residue from the original sine either as it it wouldn't constitute a sinewave even if you would continue the sections. It just looks like really weird dither. Try zooming in on the file and look what you see.

Audition in contrast produces something that looks completely random.

So I don't know what this is but it sure looks strange.

Alistair

[Nordenstam]

Hi!

Reply's been delayed due to a short vacation. Looking at clouds gets the brain going! Found the clue.

The error is in my testing procedure. It's not valid. My bad. Null testing with different bit depths simply doesn't work. Sorry for the several pages of arguments and the headaches this may have caused.

The -40dB sine wave have a certain set of sample points as it's being created in 16 bit. When nulling the 8 bit dithered file against the 16 bit file, it's the same sine wave being nulled, but it's not the exact sample points being nulled! To null the same sample points, the sine wave needs to be generated in an 8 bit dithered or truncated system, or converted from >8 bit to 8 bit with truncation distortion or dither noise. There's no avoiding this fact, it's the law of quantization. To mix 16 and 8 bit word length files as equal starting points for the test is the logical error.

There is indeed a copy of the low level information in the noise! I should have tested with a -90dB sine wave dithered&truncated to 8 bit and mixed with the 32 bit original, it's SO obvious at that level..




The choice of -40dB sine and three more at succeedingly lower levels meant that this fact was exaggerated to the point where there was obviously something wrong.

Thank you all for the persistence in pointing out that there is something iffy in the residual file. Despite my way inappropriately insistent arguments! Much appreciated.


All the best,

Andreas Nordenstam

[monomer]

Originally Posted by Lupo

Hi!

Reply's been delayed due to a short vacation. Looking at clouds gets the brain going! Found the clue.

The error is in my testing procedure. It's not valid. My bad. Null testing with different bit depths simply doesn't work. Sorry for the several pages of arguments and the headaches this may have caused.

The -40dB sine wave have a certain set of sample points as it's being created in 16 bit. When nulling the 8 bit dithered file against the 16 bit file, it's the same sine wave being nulled, but it's not the exact sample points being nulled! To null the same sample points, the sine wave needs to be generated in an 8 bit dithered or truncated system, or converted from >8 bit to 8 bit with truncation distortion or dither noise. There's no avoiding this fact, it's the law of quantization. To mix 16 and 8 bit word length files as equal starting points for the test is the logical error.

There is indeed a copy of the low level information in the noise! I should have tested with a -90dB sine wave dithered&truncated to 8 bit and mixed with the 32 bit original, it's SO obvious at that level..

Attachment 121857


The choice of -40dB sine and three more at succeedingly lower levels meant that this fact was exaggerated to the point where there was obviously something wrong.

Thank you all for the persistence in pointing out that there is something iffy in the residual file. Despite my way inappropriately insistent arguments! Much appreciated.


All the best,

Andreas Nordenstam

LOL.,., the NIGHTMARES it caused me.