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

[lfm]

- Is your position that the waveforms in music at 16 bits are 100% identical to the same waveforms at 24 bits, with 0 deviation of any kind?

i can answer that.

yes, 16-bit and 24-bit sound are bit-for-bit identical up to 16 bits of resolution. the differences only start manifesting whenever you go beyond the 16-bit range.

I don't get your reasoning, sorry. You will compare full 0 dBFs 16-bit to listening to -30 dBFs 24-bit - how does this compare or matter for any purpose?

If using -30 dBFs peaks levels, 24-bit(reality 20 bits directly from adc) - would use the same amount of bits as 0 dBFs 16 bit recording.

This is one of the main reasons to record at 24-bit, you don't loose out so much recording well within the 0 to -30dBFs level while you are forced not to loose resolution to keep up close to 0 dBFs for 16 bit.

That about recording.

-30 dBFs peak listening level - it's like musak levels in stores, you barely hear it. For 16-bit this is almost square wave you listen to, +/- 500 steps.

So if we say that SPL 85 dB is comfortable normal listening, that means 55 dB SPL which is whispering.

One reason people hate remote controls that do digital volume reduction is - since quality degrade to something not nice. Listening to cd this is +/- 500 steps to represent to highest volume at that low level.

I had one dvd player that did this, unusable to use volume remote.

So compare -30 dBFs both on 16 bit and 24 bit - that is +/- 512 steps compared to +/- 32767 steps. Even more obvious how low signal levels like on a lot of classical music will sound on 16-bit compared to 24-bit.

So don't compare only full loudness war EDM level - think about dynamics and everything.

How do you think violins from orchestra sound at these low levels 16-bit?
+/- 500 steps or 32767 steps 24-bit????

Of course you loose musical content by going 16 bit.
And theory is that this is what vinyl is doing better and part reason for it's return.

So compare at proper listening levels and 20-bit is capable of producing 16 more steps for every step 16 bits from a AD converter. If processed signal at full 24-bit or better this is 64 steps each bit for 16 bit.

Some claim then - this only make noise floor below what is silent anyway.
I'm saying you also loose musical content at 16 bit.

And intend to find out if any difference that is audible. Recording 24/96 and listen to that, render to 16/44 and listen to that, and recording at 16/44 from start and listen to that.

I also have these test patterns from RMAA that I used to look at SRC from Cubase and other daws. Maybe numbers tell a story too. You can't really listen to these frequency sweeps. But will involve analog loopback from one dac to AD on interface again.

[mcbpete]

Your theory holds up in the theoretical world where speaker cones and headphone drivers move in discrete steps, however in the psychical world such digital stepping doesn't happen at such sample rates.

[Ploki]

So compare at proper listening levels and 20-bit is capable of producing 16 more steps for every step 16 bits from a AD converter. If processed signal at full 24-bit or better this is 64 steps each bit for 16 bit.


I'm saying you also loose musical content at 16 bit.

And intend to find out if any difference that is audible. Recording 24/96 and listen to that, render to 16/44 and listen to that, and recording at 16/44 from start and listen to that.

I also have these test patterns from RMAA that I used to look at SRC from Cubase and other daws. Maybe numbers tell a story too. You can't really listen to these frequency sweeps. But will involve analog loopback from one dac to AD on interface again.

i will say it again - this is not how it works AT ALL. What you are saying is wrong on a completely theoretical level.

You shouldn't think about decibels in terms of digital audio at all, because it's not computed in decibels, it's computed in linear scale and when you start thinking in binaries and linear scale it makes absolutely more sense.

if you have orchestra at -30dBFS you are effectively using 19bits of resolution before you hit ADC, after you hit ADC you are using 15bits on the worlds best converters.
So you loose much more than you would with a dithered 16bit file normalized to 0dBFS...

why did you ignore everything i replied directly to you? is it inconvenient because it goes 100% against your narrative?

as far as digital level controls go: yes, if they work in integer. If they don't, they're fine.

And theory is that this is what vinyl is doing better and part reason for it's return.

it's doing better because of hipsters. roughly half of vinyl buyers don;t even own a turntable...

[Burillo]

I think this is another point that brings confusion. When we upsample to 64 bits for mixing we are basically just putting the same data in a larger container. It's like putting a cup of water in a gallon bucket. We still only have the equivalent of a cup of water.

yes! exactly! so, a waveform sampled in 16-bit is bit-for-bit identical to one sampled in 24-bit, except for the -96 to -144 dB additional range (i.e. lower noise floor).

However, there is a difference past the initial conversion.

but that is not what anyone was talking about. no one was saying recording and processing in 16-bits is the same as recording in 24-bits and processing in 64-bits. we're talking about playback here. and for playback, 16-bit format is bit-for-bit identical to 24-bit, bar the -96 to -144 dB range.

note the distinction: not "more detail", but "less noise". we're talking about noise that's 96dB (~110+dB with dithering!) lower than the signal. there is no way that difference is audible - frequency masking, dithering and Fletcher-Munson curve takes care of that. i'm sorry, but if you claim you hear this difference - you're plain wrong, and you're imagining things.

again, i invite you to take any 24-bit recording or even a full mix, crush it to 16-bit, and see for yourself the difference it makes (or rather lack thereof).

At the end, when you export at any bit depth you are making a choice on how much detail to preserve / how much quantization error to endure, which is what I call the "nicks and scratches" that we put on a high res wave when it is being stuffed into a limited container like 16 bits (or any other bit depth, they all have a limit, they all introduce their own amount of noise and error).

again, you're making an emotional argument and are using loaded language like "being stuffed into a limited container". some limits matter, others don't. 16-bit format is not "limited" in any audible way - it provides 96dB of dynamic range in the worst case scenario, often bigger than that. when you're playing music back, you're not using even half of that dynamic range, let alone the entirety of it, so the fact that the format has limits does not in any way imply that using that format audibly degrades quality.

I don't get your reasoning, sorry. You will compare full 0 dBFs 16-bit to listening to -30 dBFs 24-bit - how does this compare or matter for any purpose?

i'm not sure i get your point. nowhere did i mention comparing 0dBFS 16-bit to -30dBFS 24-bit.

If using -30 dBFs peaks levels, 24-bit(reality 20 bits directly from adc) - would use the same amount of bits as 0 dBFs 16 bit recording.

if you mean 0dBFS 16-bit recording uses the same amount of dynamic range that a -48dB dBFS 24-bit sound would use, then yes, it will. however, that's not the point. the point is, given two 0dBFS sounds, one 16-bit and one 24-bit, the entire range from 0 to -96dB will be bit-for-bit identical for both 16-bit and 24-bit sounds.

This is one of the main reasons to record at 24-bit, you don't loose out so much recording well within the 0 to -30dBFs level while you are forced not to loose resolution to keep up close to 0 dBFs for 16 bit.

why do you keep going back to recording? no one says recording in 16-bit is the way to go - everyone agrees with 24-bit recording.

-30 dBFs peak listening level - it's like musak levels in stores, you barely hear it. For 16-bit this is almost square wave you listen to, +/- 500 steps.

OK i lost the thread here. what on earth are you talking about?!

So don't compare only full loudness war EDM level - think about dynamics and everything.

your weird "-30dBFS" tangent aside, i am "thinking about dynamics and everything". 16-bit sound provides 96dB of dynamic range worst case scenario, ~110+dB with dithering. do you seriously think that there is any music that uses even half of that dynamic range?

How do you think violins from orchestra sound at these low levels 16-bit?
+/- 500 steps or 32767 steps 24-bit????

why don't you take a listen? you keep referring to "steps" and numbers, and making these passionate diatribes about how more steps is better, but have you actually performed a controlled listening test? take you favorite "muh dynamics" 24-bit violin recording, and bit-crush it to 14 bits, and try doing a blind test with MCompare. let me know how many times you've guessed correctly.

And theory is that this is what vinyl is doing better and part reason for it's return.

i already mentioned this, but it seems to have went above your head, so i'll repeat: going for higher sampling rate and bit depth will take you further from vynil, not closer to it. vinyl has less dynamic range and less frequency range than a 44.1/16 recording. that's just a physical fact.

Some claim then - this only make noise floor below what is silent anyway.
I'm saying you also loose musical content at 16 bit.

yeah, some say this, some say that. you see, reality doesn't give a shit what "some" say. there are facts. and the fact of this reality is, 16-bit file format only "loses" content below -96dB. do you hear that low, in a full mix? no, you don't. i know you don't, because that's just how our ears work, which is also a fact of this reality.

And intend to find out if any difference that is audible. Recording 24/96 and listen to that, render to 16/44 and listen to that, and recording at 16/44 from start and listen to that.

setting "recording in 16-bit" aside (we've already addressed this, yet for some reason you keep coming back to this stupid point), have you actually done this? have you taken a 24/96 mix, downsampled it to 44/16, then upsampled it back to 96/24, and compared with the original? i can guarantee that you haven't, because it's one of those things that you think are so obvious and intuitive that you don't even need to test it to know it's true, and this is why you're wrong.

[Ploki]

jesus all this religious bullshit around audio pisses me off so much

[Ploki]

Your theory holds up in the theoretical world where speaker cones and headphone drivers move in discrete steps, however in the psychical world such digital stepping doesn't happen at such sample rates.

or in a world where DACs would actually output a discrete waveform.

[kmonkey]

Once in my life i was dating real female model.

But it was just once.

I don't think he (i think it/that was she) ever noticed i was creating 16 bit songs.

So it works.

[Etienne1973]

Your theory holds up in the theoretical world where speaker cones and headphone drivers move in discrete steps, however in the psychical world such digital stepping doesn't happen at such sample rates.

or in a world where DACs would actually output a discrete waveform.

What I've learned: DACs reconstruct digital sample & hold to analogue flowing current, right?

[Ploki]

or in a world where DACs would actually output a discrete waveform.

What I've learned: DACs reconstruct digital sample & hold to analogue flowing current, right?

yep. http://www.craigmandigital.com/educatio ... s_dsd.aspx

it's easy to implement anyways, it's essentially just a filter, since you can look at "holds" as very high frequency content (square waveforms). What happens if you filter a squarewaveform to its fundamental? It becomes a sine.

[Burillo]

or in a world where DACs would actually output a discrete waveform.

What I've learned: DACs reconstruct digital sample & hold to analogue flowing current, right?

yep. http://www.craigmandigital.com/educatio ... s_dsd.aspx

it's easy to implement anyways, it's essentially just a filter, since you can look at "holds" as very high frequency content (square waveforms). What happens if you filter a squarewaveform to its fundamental? It becomes a sine.

subject matter side, the 10K square wave in 44.1 looks very bad... if you don't know what "band-limited signal" is in fact, that exact thing was discussed in the video by Ogg Vorbis devs on the very first page. whenever people point to such reconstruction and use that as evidence that "digital audio sucks", that's when you know the person has no frickin clue what they're talking about

[Ploki]

subject matter side, the 10K square wave in 44.1 looks very bad... if you don't know what "band-limited signal" is in fact, that exact thing was discussed in the video by Ogg Vorbis devs on the very first page. whenever people point to such reconstruction and use that as evidence that "digital audio sucks", that's when you know the person has no frickin clue what they're talking about

Should've played it back on a speaker system and recorded it with a microphone

[imrae]

What I've learned: DACs reconstruct digital sample & hold to analogue flowing current, right?

Correct, with a steep low-pass filter at the ideal frequency. But even if they didn't, you'd still never get perfect stair steps because speaker cones have to follow Newton's laws.

[Guenon]

Thank you for clarifying your thoughts. I'm now convinced you and me are talking about the same thing, just from different angles. You are saying the waveform is "the same" albeit hampered by a tiny bit of noise.

I'm saying that the noise is the manifestation of the nicks and scratches that you put on that waveform when you sampled it at 16 bits, and you would have less of that noise in a 24 bit system (but still some manner of faint noise). I believe you will agree with that.

To you that means that it is the same waveform, just slightly noisier. To me that means that it is the same waveform, just slightly less accurate because it is constrained to the 16 bits space. Same thing, case closed.

Hehe, "case closed." As I said already, you are on a respectable path, but at the same time you are just letting your intuition get in the way of learning how it actually works. No biggie. From this point forward, I almost might as well answer with quotes of what I've already said, anyway.

"Nicks and scratches that you put on the waveform" is an unorthodox way of describing noise, but as long as you really know you are talking about very low level hiss and decide to call it that, then hey, what ever works for you !

Just remember, in order to be consistent, you would have to use that terminology in other contexts as well. Imagine you are talking to someone who runs a top class studio and chooses to record a project on top class analog tape , and using your terminology, the relatively low level hiss in that system puts quite a bit of nicks and scratches on the waveform. For consistency's sake, you should point that out in those terms as well. In my experience, "nicking and scratching the waveform" hasn't been that much of a talking point when dealing with state of the art tape equipment, so it's sort of humorous to see the (dramatically lower level) hiss in digital systems described in such terms -- just because of the mental imagery of how the hiss is brought into existence in a digital process.

But as I said, as long as you are consistent and really know what you are talking about, then, well....

I still recommend you watch the videos I linked on the first page, though, do meaningful experiments and stay at it. On your journey, I mean. You are already in much clearer waters than lfm is, those descriptions are just so deep in pseudo.

Take care!

[jochicago]

"Nicks and scratches that you put on the waveform" is an unorthodox way of describing noise

I'm not trying to describe noise. I'm focused on describing what's happening to the waveform when you stuff it into a sampling format. I used that expression to try to use something meaningful beyond the specs of the process, and also to connect it with the idea that you guys have that "the only difference between resolutions is noise". That's not how it works. It's close enough, but not it.

Digital wave files are a sampling format. Noise is a byproduct. If we want to determine the actual difference in these formats, the conversation we should be having is about what's happening in the sampling process, not focus on the noise byproduct.

And you are right that we seems to be going a bit in circles. I think this is my last attempt at trying to express myself with clarity, if this doesn't illustrate it then I've failed at portraying it.

I made this graph as a visual representation of what happens when you grab an analog audio wave and you put it in a digital wave file.

Let's zoom in to the tiny tip of an audio peak:

wave peak.jpg

-

wave dots.jpg

I don't know that I can make it much clearer than this. This chart illustrates how the process actually works. No guesses or intuition here, this is what computers do.

1. We start with a source waveform.
2. According to the system's resolution, we take samples of the waveform. We take only as many samples and in the level of detail that the resolution calls for.
3. We then use those "dots" to reconnect the lines using smart algorithms that are attempting to reconstruct the original waveform from just the dots, because the dots info is what is stored.

It doesn't get any simpler than this:
Larger sampling rate = more dots closer in time
Larger bit depth = the dots are more precisely placed on the wave line

After the dots have been sampled, the system has to re-imagine the waveform only from the dots available. This is what a digital wave file is.

So regardless of the resolution, it will always be using a re-imagined version of the waveform that will always be missing some level of detail from the original. Even at hi-fi levels. With higher resolutions it will just get closer and closer to portraying all the nuance in the original waveform.

The change of shape from the original into a re-imagined shape is what I called "nicks and scratches" on the waveform. I meant to represent that, in practice, when you stuff a waveform into a digital format you are keeping the overarching contour of the waveform and losing any finer detail on it because you are only recording as many dots as the file resolution called for, and the gaps in between have to be imagined with approximations.

Higher resolution means more dots to be able to reconstruct the waveform more faithfully.
This is exactly how a wave file works. There are no caveats. It is not open to interpretation.

[Etienne1973]

https://youtu.be/Rc_-eavKptY

I'm confused.

1. Isn't this video the proof that the count of steps vertically (bit depth) matters at low signal levels?

2. Does this mean the higher the signal level the less bit depth matters?