Flac 16 bit or 24 bit Qobuz

[castleofargh]

I think the problem lies when some claim their opinion is science and that any dissenting opinions are not going against an opinion, but science itself..

Those instances are very likely an Appeal to Authority. A well-known logical fallacy.

I agree with that, actually. There is a general lack of caution when claiming something, and that's found on any side of any argument. But of course, some things are factually established and considered scientific knowledge. No point in presenting those as personal opinions.

What annoys me just as much as trying to make something look proved when it hasn't been, is how often people won't take it as that one person claiming something, but will instead start to oppose, the entire subsection or even Science in general. It's weird, obviously easy to do as it happens so often, but weird nonetheless. I disagree with that one guy, you know what will make it easier for me? If I start fighting an entire group or the very concept of science instead.

Why does rounding cause audible distortion in 16bit, and especially 8 bit, but not, say, in 24bit?

24bit is irrelevant because nothing can replay those LSBs. You can do whatever you want to them, they won't come out on any sound system.

16bit probably wouldn't be audible unless you go for extremely quiet signal and high playback level(to bring the LSB level somewhere you can hear it, and it won't be entirely masked by the louder music). All the times someone talked about passing such a test, he used that trick instead of typical music at typical listening level.

8bit LSB is still loud enough for a lot of the lowest signal to not be masked by louder content. But as mentioned, with noise shaping, even picking up 8 bit noise can be challenging.

Distortion or noise is just a nomenclature thing. In general, correlated errors are called distortion, and random stuff, like the wife talking while you listen to music, that's noise.
But of course that would be clear and simple, so someone introduced the concept of correlated noise, just to mess with us.

 

[71 dB]

Why does rounding cause audible distortion in 16bit, and especially 8 bit, but not, say, in 24bit?

16 bit rounding errors aren't actually audible (unless you gain ride or something), but dither is used anyway because it is so easy. Just add a little bit noise before rounding and distortion due to rounding in completely gone! 24 bit version can actually contein so much noise it acts as "self-dither" when truncating to 16 bit. So, 16 bit dither is not "needed", but it is done because it is so easy and it is mathematically the smarter way to truncate bits...

 

[71 dB]

Subjective impressions don’t require any thought at all. It’s just a feeling, and it isn’t likely that my feeling is the same as yours. So subjective impressions are unique to the individual, they don’t apply to anyone else. Science applies to anyone in the physical world we inhabit.

People should learn when to listen to their feelings. If you feel you'd like a new movie, you perhaps do while science knows nothing about how much you'd like the movie. Feelings have theirs place. Science has it's place. Applying those things correctly gives the best results. Feelings are difficult to ignore, but people should try whenever feelings aren't the thing to believe. Science doesn't know why I like the music I do, but it surely knows a lot about the planet Jupiter!

 

[71 dB]

Where does the NOISE come from when you down convert?

Let's do something even more extreme....what happens when you do 24bit versus 8bit?

You replied to my post with these questions, but I believe they have been answered to you by now.

 

[gregorio]

Why does rounding cause audible distortion in 16bit, and especially 8 bit, but not, say, in 24bit?

Maybe you don’t realise how bits work? The maximum signal amplitude (0dBFS) is the same regardless of whether you use 8, 16 or 24 bits. What changes with the bit depth is not how big/loud the signal you can encode but how small/quiet. The 8th bit is the same level regardless of whether you’re using 8, 16 or 24 bits but with 16 and 24 bit you can encode signals lower in level than the 8th bit (roughly a 6dB lower level signal for each additional bit). So rounding to 8bit is more audible than rounding to 16bit because the 8th bit is far louder/higher level than the 16th bit.

Rounding and truncation cause distortion because the resulting error is a signal that is correlated, so it periodically sums with your wanted/musical signal causing distortion, while dither forces the error into relatively benign noise which is decorrelated and therefore does not sum. However, as others have stated, even the distortion from truncating to the 16th bit will be inaudible at reasonable listening levels. At 8bit it can be audible and at 24bit it’s so low in level it cannot even be resolved into sound.

G

 

[eq1849]

Maybe you don’t realise how bits work? The maximum signal amplitude (0dBFS) is the same regardless of whether you use 8, 16 or 24 bits. What changes with the bit depth is not how big/loud the signal you can encode but how small/quiet. The 8th bit is the same level regardless of whether you’re using 8, 16 or 24 bits but with 16 and 24 bit you can encode signals lower in level than the 8th bit (roughly a 6dB lower level signal for each additional bit). So rounding to 8bit is more audible than rounding to 16bit because the 8th bit is far louder/higher level than the 16th bit.

Rounding and truncation cause distortion because the resulting error is a signal that is correlated, so it periodically sums with your wanted/musical signal causing distortion, while dither forces the error into relatively benign noise which is decorrelated and therefore does not sum. However, as others have stated, even the distortion from truncating to the 16th bit will be inaudible at reasonable listening levels. At 8bit it can be audible and at 24bit it’s so low in level it cannot even be resolved into sound.

G

To compare apples to apples, let's look at just the TOP bit of 8,16,24,32 bit audio...the -6db to 0db range.

From my understanding...

...the number of Quantitation Steps within the TOP bit of each:

32bit ~2.1 Billion
24bit ~8.3 Million
16bit 32,768
8bit 128

If that's the case, PRECISION is greater with more STEPS over the given range.

Quantization noise, ultimately, is a byproduct from lack of PRECISION.

And the greater the bit depth, the more PRECISION, along with greater Dynamic Range and a lower Noise Floor.

If I'm understanding all of this correctly, what is the mathematical justification that I would ever want to drop from 32 or 24 bit audio to 16 bit, unless I have a format constraint?

 

[blaked]

If I'm understanding all of this correctly, what is the mathematical justification that I would ever want to drop from 32 or 24 bit audio to 16 bit, unless I have a format constraint?

Why not 64-bit? Why not 128-bit? Why not 2048-bit?

If you're throwing darts at a dart board, do you score any fewer points if you land in the 100-point circle but you're just inside the edge of it instead of dead center? At some point you can just say that you're "precise enough".

That's what has been pointed out here repeatedly. Unless you're gain riding quiet passages, 16-bit is easily "precise enough" because you won't hear the dither noise. Nobody's stopping you if you want to keep 24-bit files, but for listening purposes you're not doing much but taking up extra storage space and/or bandwidth. (On the plus side, you're not really hurting anything, either, beyond those storage and bandwidth factors.)

 

[eq1849]

Why not 64-bit? Why not 128-bit? Why not 2048-bit?

If you're throwing darts at a dart board, do you score any fewer points if you land in the 100-point circle but you're just inside the edge of it instead of dead center? At some point you can just say that you're "precise enough".

That's what has been pointed out here repeatedly. Unless you're gain riding quiet passages, 16-bit is easily "precise enough" because you won't hear the dither nosie. Nobody's stopping you if you want to keep 24-bit files, but for listening purposes you're not doing much but taking up extra storage space and/or bandwidth.

How many amplitude "steps" are on this function between 1 to -1?

Hint, it doesn't top out at 65,536.

https://www.wikihow.com/Graph-Sine-and-Cosine-Functions

 

[castleofargh]

Why would I ever want to drop from 32 or 24 bit audio to 16 bit, unless I have a format constraint?

You do what you want. Nobody will stop you.

I would argue that none of my favorite albums have 16bit truncation as the highest noise on the 16bit files. So what extra "PRECISION" would I get from a 24 or 32bit file of those albums? More accurate noises? More noises?

Obviously I don't care for more than 16bit because as I already explained, I don't hear a difference with more bits(in an ABX with the 24bit converted to 16 then back to 24). At the risk of losing my hardcore objectivist membership card, if I don't hear it in my music, I don't really care for it.

 

[sander99]

How many amplitude "steps" are on this function between 1 to -1?

Hint, it doesn't top out at 65,536.

https://www.wikihow.com/Graph-Sine-and-Cosine-Functions

This picture of course is meant to depict a continuous sine function that goes through all real numbers in the interval [-1, 1], hence uncountibly infinite many values (and without steps).
Although of course it is not really a continuous line in that actual digital image because that image consists of independent pixels.
Very unlike the output of a proper DAC with correct reconstruction filtering, that actually outputs a real continuous signal with uncountibly infinite many values (without steps).
Now for any 32 bit signal, also a "big" signal at maximum level, the following is the case, after conversion to analog:
-The difference between that 32 bit signal and that signal truncated to 20 bits is a difference signal (the one substracted from the other) with a maximum level at -120 dB.
-The difference between that 32 bit signal and that signal truncated to 16 bits is a difference signal (the one substracted from the other) with a maximum level at -96 dB.
In all these cases the output of the DAC is still/again a real continuous signal without steps.
And the differences are to small to hear under normal listening conditions (without gain staging tricks, gain riding etc.). Although the truncated to 16 bits version may be on the edge (not in any practical scenario though), luckely proper downconversion to 16 bits with proper dithering will be far on the save side of audibly transparent again, like the truncated 20 bits version.

 

[blaked]

How many amplitude "steps" are on this function between 1 to -1?

Hint, it doesn't top out at 65,536.

That might be relevant if we were in a parallel universe where digital audio worked completely differently than it does.

You're completely missing that the number of bits affects absolutely nothing but the depth of the noise floor. The analog signal is reproduced with 100% accuracy regardless (provided Nyquist-Shannon conditions are met, of course).

 

[eq1849]

The analog signal is reproduced with 100% accuracy regardless (provided Nyquist-Shannon conditions are met, of course).

You need infinite perfection to make it happen, no?

how many "steps" of amplitude do you think are between 1 and -1 on a sin(x) function?

 

[XTF1]

16 bit rounding errors aren't actually audible (unless you gain ride or something), but dither is used anyway because it is so easy. Just add a little bit noise before rounding and distortion due to rounding in completely gone! 24 bit version can actually contein so much noise it acts as "self-dither" when truncating to 16 bit. So, 16 bit dither is not "needed", but it is done because it is so easy and it is mathematically the smarter way to truncate bits...

I work in an industry where we use a lot of realtime sensors feeding digital-based controls and models-based controls. The very same science-based concepts apply… For example, dithering is used on all the sensors.

The difference is… we don’t have to deal with a “sensor-phile” crew (audiophile-equivent) arguing that our control models are wrong because of quantization errors, and how we deal with these…

 

[sander99]

You need infinite perfection to make it happen, no?

how many "steps" of amplitude do you think are between 1 and -1 on a sin(x) function?

Two things:
1. The DAC needs only a few samples per cycle to reconstruct a continuous sine wave with uncountably infinite many values. If the samples do not have the "infinitely precise" correct value then there is added noise. But the result is still a continuous output (the sine wave plus noise). So no, we don't need infinete perfection to make it happen.
2. I admit it may come over as a bit philosophical or a matter of semantics but what @blaked means is that even if there is noise+signal, the signal is still in there, with less bits you get (more noise)+signal, but the signal is still in there... Anyway, I prefer the noise to be inaudible, luckely with 16 bits and dithering that is no problem at all.
(See the video - also linked in bishot's sig - "The truth about bit-depth and digital audio resolution" : )

 

[blaked]

You need infinite perfection to make it happen, no?

No. I think maybe you're mixing up concepts. An analog signal can be reproduced with 100% accuracy provided the signal is band limited and sampling is done at a frequency at least twice that of the highest frequency in the signal (i.e., Nyquist-Shannon), regardless of the number of bits used for each sample.

People commonly picture a stair-step reproduction of the analog signal, but that's not what happens. Plenty of explainers are guilty of perpetuating that, but it's an understandable gaffe when trying to conceptualize it. It's more accurate to think of disconnected pinpoints which a DAC can use to reconstruct the analog signal by drawing a line through the dots. Note that these are not just straight lines, like a connect-the-dots picture. A calculation is done by the DAC to find the one and only one path that mathematically meets the appropriate conditions to flow through the series of dots. I'm oversimplifying, but that's basically what happens. It's why band limiting is necessary for Nyquist-Shannon, too. If the signal wasn't band limited, there could be more than one path that meets the conditions, and the DAC could choose the wrong one (aliasing, which is bad). Band limiting ensures that the DAC will come up with the right answer every time, even if the number of bits per sample is small.

Where the number of bits comes in is in the rounding error. Looking at one sample (one "slice" of the analog signal in time), the record of where the line passed through can be recorded with as little as one bit (0 for one level vs. 1 for the other level). Fewer bits means the approximation of where the line passed is less precise, but this isn't actually a big problem for reconstructing the signal itself. If we've met our Nyquist-Shannon conditions, we're still going to reproduce our analog signal accurately.

However, a small number of bits means greater rounding error, which means more noise/distortion. If we apply dither so that the noise is uncorrelated with the signal, our low-bit sample will still be pretty unlistenable, but not because the signal wasn't reproduced accurately! Even with a small number of bits per sample, there's still only one path through the dots that works. It'll be unlistenable because the level of the uncorrelated noise is so high! Increasing the number of bits allows better approximations of where the dots are, i.e., less rounding error, which allows for less noise. Once you get to about 12 or 13 bits, you've reduced the noise to the point where it's inaudible under most typical listening conditions. 16-bit sampling (for listening, not mixing/production) turns out to have been a very forward-thinking choice.

Edit: It occurs to me that my example of sampling with only one bit doesn’t quite work because you couldn’t dither it. Randomizing the least significant bit would eliminate the signal and replace it with white noise altogether if you only had one bit to start with! A 2-bit sample (four possible values) might’ve been a better simplification. Oh well. Either way, fewer bits just means more noise added to our accurately reconstructed signal, nothing else.