The truth about bit-depth (and digital audio ‘resolution’)

[Guenon]

I'll continue to track the thread and read any links, but I'm comfortable in my own understanding, and personally don't have more to add.

That sounds good. I'll just leave you with this thought: the very first thing you said in this thread was "I'm not an expert by any means, and definitely have spent no time trying to compare the sound of 16 vs 24 bits." Yet it didn't take long until you had, by interpreting technical documents on your own in the span of a couple of days, convinced yourself that you knew the implications better than for example DSP professionals such as Vorbis codec author Chris Montgomery (whose concise demonstrations are linked multiple times in this thread) and other experts who have spent years professionally on the subject. All this in a couple of days, from a "not an expert by any means" who has spent "no time trying to compare" the very things the discussion was about.

This isn't a put-down, believe it or not. Not even when I say this: I know the Dunning-Kruger effect is flaunted online as an easy insult, ready to go when ever someone offers a differing opinion on anything that can be portrayed in such light. So much so that even mentioning it by its name... just makes a discussion go sour. However, the conversation seems to be peacefully over, and I'll finally keep quiet as well -- and in cases like this, it's good to recognize the actual effect (not the insult kind) playing tricks on someone getting to know a new thing. I'm no exception in this regard, it goes without saying.

Just ask yourself, what are the objective odds that concepts outlined in this thread (that you call "repeating the same wrong concepts"), demonstrated to you by veteran DSP professionals like Chris (Monty), are the wrong ones -- and the ones you come up with in a day or two are the correct ones?

The problem wasn't your sources, and you weren't even that far off; as you hopefully saw, we agree on most things. It was just your intuition based conclusions, at this moment in time, on what the actual implications then are. I'm sure the conclusions will keep getting more accurate and proportioned and serve you well.

[Guenon]

my advise would be to read on about how sound is represented in binary in a PCM format. understanding this is crucial to understanding what different bit depths will give you (and what they can't give you).

[Ploki]

my advise would be to read on about how sound is represented in binary in a PCM format. understanding this is crucial to understanding what different bit depths will give you (and what they can't give you).

this!

Decibels and decimal are terrible to grasp concepts of digital audio.
Also, Nika Aldrich - Digital Audio Explained.
It's written incredibly well and even if you're not super tech savvy the concepts are presented in a manner that's easy to grasp.

[vurt]

if a tree falls in the forest, should i record using a higher bitrate or given the difference, will no one hear it?

[Ploki]

if a tree falls in the forest, should i record using a higher bitrate or given the difference, will no one hear it?

DSD for trees is best

[mcbpete]

if a tree falls in the forest, should i record using a higher bitrate or given the difference, will no one hear it?

You definitely wont if you use float.

[jochicago]

I'm a trained software engineer. Being able to study technical specs quickly is part of my day job.

I don't know DSP programming, and that's not something that can be learnt overnight. But studying the spirit of analog-to-digital, wave file encoding and reproduction, and the overall pitfalls of these technology can be done in a day if you have the skills for it.

Plenty of this is pretty simple. In a digital sampling system you have to take measurements of things and then record those measures. You have to do it at enough resolution for accurate reproduction, then you need a definition for what "accurate enough" is, since in every instance you'll be dropping fidelity regardless of what you do. That's hardly new to me or any programmer. That's how everything works, from wave files to DSLR image sensor -> jpegs. Anything that converts infinite data into a portable digital format will lose fidelity according to the choices made in the format.

Then, even beyond the limitations of resolution of the chosen format, every system has built in errors, from inevitable things like timing variability in hardware to many points of unwanted distortion. In digital systems we usually cope with these things by using correction algorithms and oversampling. The specifics of each field are unique, but every digital system has this in any field.

In terms of this conversation, I spent time studying the logic behind typical DSP correction and oversampling of digital waveforms, to understand the depths required (and currently very much in use) across the field. That's usually a very good place to focus because it quickly reveals where distortions appear, where the basic theory fails to hold steady, where the fidelity is lost and how we are coping to solve those issues. Critical things to understand when you are choosing a sampling format.

I don't want to stretch this out but I also don't want you to leave with the mistaken impression that you are right on all your points. You are most definitely not. You keep defining your understanding on the basic theorems of it, but practice is a different universe. Not just about what "humans can hear" in these formats which is already an important point to be argued, or your understanding of the rules of digital systems. But also of great importance is that most well-designed systems oversample to reduce discrepancies/system distortion, and they do so for very good reasons.

You guys need to do a lot more learning about how these digital systems work in practice, why you'd want to be at 24 bits for capture and storage, and why 44.1k is simply not a great sampling rate in 2019.

Notice I avoided going into specific posts here or pointing at individual people. No need to reply to me/this post. You already said your parts repeatedly, and I'm not arguing any more points. I only replied because you quoted me again.

[Burillo]

you can refer to oversampling and other complex technical stuff all you want, and we can spend infinity over arguing how good is good enough...

...but at the end of the day you still claim (and apparently continue to do so) that, within dynamic range of 16-bit sound, 24-bit format somehow reproduces things more faithfully. that is just factually incorrect due to how PCM encoding works, and has been brought up to you a number of times already, but has fallen on deaf ears.

(by the way, you're not the only software engineer here, and some of us actually have worked with DSP before... but i digress)

[Guenon]

I'm a trained software engineer. Being able to study technical specs quickly is part of my day job.

I don't know DSP programming, and that's not something that can be learnt overnight. But studying the spirit of analog-to-digital, wave file encoding and reproduction, and the overall pitfalls of these technology can be done in a day if you have the skills for it.

Well, in that case...

May I refer you to my post above

Notice I avoided going into specific posts here or pointing at individual people. No need to reply to me/this post. You already said your parts repeatedly, and I'm not arguing any more points. I only replied because you quoted me again.

Ah yes, the old "I'll say I'm quiet and then I'll reply anway when the other person says something" game. Hehe. Threads like this are notorious for making people play that one. Look at how good I am at it, too. Will you reply again? Will I?

[vurt]

i wont.

[vurt]

damn!

[Etienne1973]

Does binary keeps information partially intact when truncated instead decimal looses info completely?

For Example:

24Bit word truncated to 16Bit

1011 1111 0101 1101 1010 0110
____ ____ 0101 1101 1010 0110

decimal number truncated

1'200'377
120'037

[Burillo]

Does binary keeps information partially intact when truncated instead decimal looses info completely?

For Example:

24Bit word truncated to 16Bit

1011 1111 0101 1101 1010 0110
____ ____ 0101 1101 1010 0110

decimal number truncated

1'200'377
120'037

"truncated decimal" would rather be 200377, not 120037. truncation starts from the top.

[Ploki]

i don't its possible to truncate an integer.
truncation of decimal numbers (as opposed to rounding) usually happens right of the floating point

[Burillo]

i don't its possible to truncate an integer.
truncation of decimal numbers (as opposed to rounding) usually happens right of the floating point

anything can be truncated.

"truncation" is a process of fitting a bigger depth number into a smaller one, dropping the extra data. for binary, depth is in bits (i.e. powers of two). for decimals (as in, not binary numbers) - it's powers of ten.

floating point truncation is a different thing entirely though.