Bit depth revisited

[jayfrigo]

Potentially of interest... Not a perfect example, but worth thinking about. The unfiltered, combined square wave doesn't appear to have an accurate amplitude reading at any single given point of the resulting triangle wave.

[okydoky]

Originally Posted by Audiop

But aren't you describing one single phenomena as two? What you mention here is the same thing, no?


/Peter

hi,

not really. perhaps "two aspects of the same phenomena" would make sense. but they are severable aspects, even if related.

i understand what you are saying [i think ]. however, the discrete value that the quantizer assigns to the sample, and the magnitude of the noise [which is admittedly the remainder of the initial analog value minus the value that the quantizer assigns to the sample] can be separately identified and given separate values.

as i believe ozpeter's adobe audition chart showed, the sample values, even if they contain a noise component, are different between the two bit depths [at least i think that is what he was displaying]. the 24 bit samples are more accurately, or "finely", represented. a 16 bit quantizer cannot represent those values at all.

i do see the concept that if you remove the error, you must be left with the signal, so its not like i don't see the issue.

with dither, you can remove the quantization / truncation noise of the 16 bit audio below -96dBfs. but it seems to me that it does not "rewrite" all the amplitudes to the same values that are present in the 24 bit file. those values are unrepresentable in 16 bits. and with regard to the higher amplitudes , it seems to me that there is more accuracy represented in the 24 bit samples than could be represented in the 16 bit samples.

dithering to 16 bits does not give you the same amplitude values for the waveforms as you had when the file was 24 bit. the values of the 24 bit samples will be more accurate, with less quantization error [even if only "slightly"].




right.

[okydoky]

Originally Posted by jayfrigo

In the end, you put in a signal, and you get the same signal out, only noisier, within the limits of the system, i.e. nothing above Nyquist or below LSB. You're getting the individual pieces correct, but seem to be missing the application of them as a whole, how the entire system works together sample to sample, which results in a linear system (within limits previously stated).

It really is too late to continue. Too bad we've scared Dan Lavry off. It's time to tag somebody like him or Dave Collins or Daniel Weiss so I can get some rest.

hi,

i hear you. but it seems to me that the noise [quantization error] is directly related to the lack of accuracy, and that it can be separately identified, and quantified.

my research into this [admittedly a while ago] showed that the removal of truncation / quantization noise by use of dither in bit rate reduction [or other conversion], did not simultaneously perfect the amplitude levels of the individual samples.

granted, the resulting wave form will not be grossly different as between the 24 bit and 16 bit samples, but it is not possible for 16 bits to represent all of the values that 24 bits can represent, so there will be some "rounding" of the individual samples.

at higher amplitudes, the lsb is functioning as part of a larger word, and it is not susceptible to the problem with correlation, no?

are you saying that the dither supplies enough random signal, and somehow it can be filtered to change the error signal into accurate signal?

or are you saying that the dither signal bits somehow get appended to all the digital words of the higher amplitude signals also? hmmm. but the dither is statistically random, and there is nothing to suggest that the original signal was random. so you would get a more finely "resolved" signal, but not necessarily the same signal as the input signal [more finely resolved, but not necessarily accurate]. that's kind of where i'm going with the "statistical improvement" [but it more or less conjecture]. i understand that there has to be some time domain to bring statistics in, but the samples could still be considered individually.

however, the signal is already represented as accurately as possible by the given bit depth when considering the individual samples.

the idea of statistical "infinite" bit depth would seem to only have relevance, if any at all, during initial conversion , because once the signal has been quantized you no longer have a continuous source. so sample rate conversion from one bit depth to a lower bit depth ["redithering"] would not necessarily result in as accurate a result.

here is a "wikipedia quote", with a number of references given there:

"Note that dither can only increase the resolution of a sampler, it cannot improve the linearity, and thus accuracy does not necessarily improve."


anyhow, daniel would be the one for this stuff, hands down i think.
weiss is in fact the sample rate conversion meister.

by the way, i am not saying that i could not be wrong about something, and i am always willing to learn. nice hearing from you.


right.

[okydoky]

Originally Posted by jayfrigo

In the end, you put in a signal, and you get the same signal out, only noisier, within the limits of the system,

hi,

and the limits of a 16 bit system are different than the limits of a 24 bit system. so the signal you get out of a 24 bit system is slightly "better" [more accurately resolved], and slightly less noisy.



right.

[jayfrigo]

Originally Posted by oky****

anyhow, daniel would be the one for this stuff, hands down i think.
weiss is in fact the sample rate conversion meister.

Yeah, he's great when it comes to digital stuff. It's too bad he doesn't do more online discussion. I've only seen him on rare occasion online.

[chris319]

Enough theory. In my own work recording voice-overs and a symphony orchestra I've standardized on 24 bits. I give myself a good 20 dB of headroom and digitally boost levels after the fact. I have oodles of headroom, never have a clip, have total control over levels and the results sound great. I do this knowing that even if I used 48 dB of headroom, the 48 dB boost would make it equivalent to a 16-bit recording.

I have written a program which creates a histogram of all samples between -20 dBFs and clip (0 dBFS), so I can see the distribution of amplitudes of the samples. Sometimes, if there is a very small number of samples at the highest recorded levels, I might apply more gain than normalization would and actually allow those samples to CLIP (Gasp!) on the assumption that they are transient peaks too short for the ear to pick up. If anyone is interested in a copy of the program I might pretty it up and make it available.

[Ozpeter]

I'm pretty sure that Audition has such a histogram within its analysis window.

Sometimes, if there is a very small number of samples at the highest recorded levels, I might apply more gain than normalization would and actually allow those samples to CLIP (Gasp!) on the assumption that they are transient peaks too short for the ear to pick up.

Gulp... I think I would recommend the Classic Hard Limiter vst, or one of many others, for that task (I mention that one because inversion tests show that it doesn't monkey with samples that it should leave alone - and it only has one knob, which suits my brain nicely).

[chris319]

I'm pretty sure that Audition has such a histogram within its analysis window.

Yes but my program is $400US cheaper

It's actually the Classic Master Limiter. I'll have a look at it.

[Audiop]

Originally Posted by chris319

I do this knowing that even if I used 48 dB of headroom, the 48 dB boost would make it equivalent to a 16-bit recording.

A 24bit recording doesn't have 48dB more SNR as compared to 16bit but typically something like 20-35dB depending on which actual designs we are looking at.

A 16bit roundtrip may have 90dB SNR and a 24bit roundtrip 115dB.


/Peter

[d_fu]

Originally Posted by d_fu

I did a test with an audio signal at -70 dBFS or so, using various dither types and no dither at all. It's interesting to hear how levels drop into quantization noise without dither, and how levels that are nominally below -96 dBFS can still be represented with good dither. Will dig up the files if there is interest... But that's OT here, material for a new thread maybe.

Done, as part of David Rick's dither thread.


Daniel

[chris319]

A 24bit recording doesn't have 48dB more SNR as compared to 16bit

Correct, but it starts out with 2^256 more quantization levels.

[okydoky]

Originally Posted by chris319

Enough theory. In my own work recording voice-overs and a symphony orchestra I've standardized on 24 bits. I give myself a good 20 dB of headroom and digitally boost levels after the fact. I have oodles of headroom, never have a clip, have total control over levels and the results sound great. I do this knowing that even if I used 48 dB of headroom, the 48 dB boost would make it equivalent to a 16-bit recording.

I have written a program which creates a histogram of all samples between -20 dBFs and clip (0 dBFS), so I can see the distribution of amplitudes of the samples. Sometimes, if there is a very small number of samples at the highest recorded levels, I might apply more gain than normalization would and actually allow those samples to CLIP (Gasp!) on the assumption that they are transient peaks too short for the ear to pick up. If anyone is interested in a copy of the program I might pretty it up and make it available.

hi,

what daw are you using to do the "normalization"? you know that gain changes after the file has been written in conversion result in further errors, right [unless they are done at a higher bit depth than the recording]?

the reason i ask about the daw is because pro tools, for example, uses a 48 bit mixer [56 bit accumulator] to handle any such level changes without mangling the audio. you record the audio at 24 bits, it gets sent to a 48 bit environment for an level / pan changes or other dsp, and then back to the bus at 24 bits.

you may already be aware of all this, and i hope i am not being redundant.

do you find that you need all that headroom, or is it just a standard that you have adopted? not trying to be "contentious" at all. i am really wondering if the typical dynamic range in what you are doing requires that kind of leeway.

by the way, i don't doubt that the stuff sounds good.


right.

[d_fu]

Originally Posted by chris319

Correct, but it starts out with 2^256 more quantization levels.

No, only 256 times as many... 16777216 divided by 65536 is 256.

2^256 is 1,1579208923731619542357098500869e+77

[chris319]

Originally Posted by d_fu

No, only 256 times as many... 16777216 divided by 65536 is 256.

2^256 is 1,1579208923731619542357098500869e+77

Bah. That should have been 2^8.

In the future please go by what I mean, not by what I say. Thank you.

[chris319]

Here is my method for recording bagpipes. Record a single monaural channel of the bagpipes. Copy that channel to a second channel, inverting the recording in the process so that + is - and - is +. Now mix the second channel back to the first and delete the second channel. The result will be the most beautiful bagpipe music ever recorded.

This works at all bit depths and sampling frequencies, BTW, provided the two channels have identical bit depths and fs.

[okydoky]

Originally Posted by chris319

Here is my method for recording bagpipes. Record a single monaural channel of the bagpipes. Copy that channel to a second channel, inverting the recording in the process so that + is - and - is +. Now mix the second channel back to the first and delete the second channel. The result will be the most beautiful bagpipe music ever recorded.

This works at all bit depths and sampling frequencies, BTW, provided the two channels have identical bit depths and fs.

hi,

superb! i never thought of that. perhaps all bagpipe music should be recorded that way. discuss......


right.

[okydoky]

Originally Posted by chris319

If anyone is interested in a copy of the program I might pretty it up and make it available.

hi,

just how pretty are we talking about?

i mean we'd have to be talkin' about one charming mother****ing program. i mean she'd have to be ten times more charming than that arnold program from green acres, you know what i'm sayin'? [jules: pulp fiction]


right.

[okydoky]

Originally Posted by jayfrigo

Yeah, he's great when it comes to digital stuff. It's too bad he doesn't do more online discussion. I've only seen him on rare occasion online.

hi,

mmm hmm. he has no axe to grind. weiss makes amazing stuff for pretty much any / all formats. they seem to even be doing dsd and dxd capable stuff now.

i think daniel weiss is actually quite accessible, and a nice guy. he might be willing to field a couple of questions if we need him.

mabe it would be good if we could distill it down a little first, though huh? i don't think it would be too productive to hit him with a big discussion on whether or not 24 bit audio is worthwhile because someone doesn't think it sounds any different.

he really knows sample rate conversion and dsp, and as soon as we figure out what the question actually is, maybe i could ask him to answer it.


right.