[StanD]

A bove ante, ab asino retro, a stulto undique caveto

Donkeys are smart enough not to fall for hirez marketing hype.

 

[Cold Train]

Donkeys are smart enough not to fall for hirez marketing hype.

MP3 is still the king of my collection, so I don't feel that I'm directly targeted by your statement

 

[StanD]

MP3 is still the king of my collection, so I don't feel that I'm directly targeted by your statement

I stream 320k (more than enough) MP3 from my Google All Music subscription and m very satisfied by the result. What's wrong with donkeys?

 

[Cold Train]

I stream 320k (more than enough) MP3 from my Google All Music subscription and m very satisfied by the result. What's wrong with donkeys?

I personally believe that everything above 16/44 is a scam !

'' Even with a $4000 Focal Utopia on it's head, a donkey remains a donkey,''

 

[StanD]

I personally believe that everything above 16/44 is a scam !

'' Even with a $4000 Focal Utopia on it's head, a donkey remains a donkey,''

You shouldn't pick on donkeys, they're as smart as any audiophile. They're stubborn because they are smart, ask one and they'll tell you, "You want tricks, get a dog."

 

[Cold Train]

You shouldn't pick on donkeys, they're as smart as any audiophile. They're stubborn because they are smart, ask one and they'll tell you, "You want tricks, get a dog."

a cane muto et aqua silente cave tibi

 

[StanD]

a cane muto et aqua silente cave tibi

My dog drinks bottled water. Hold it, I don't have a dog, but I do have headphones.

 

[Cold Train]

My dog drinks bottled water. Hold it, I don't have a dog, but I do have headphones.

 

[Cutestudio]

Dither: Essentially during the conversion process a very small amount of white noise is added to the signal, this has the effect of completely randomising the quantisation errors. Randomisation in digital audio, once converted back to analogue is heard as pure white (un-correlated) noise. The result is that we have an absolutely perfect measurement of the waveform (2*) plus some noise. In other words, by dithering, all the measurement errors have been converted to noise. (3*).

Hopefully you're still with me, because we can now go on to precisely what happens with bit depth. Going back to the above, when we add a 'bit' of data we double the number of values available and therefore halve the number of quantisation errors. If we halve the number of quantisation errors, the result (after dithering) is a perfect waveform with halve the amount of noise. To phrase this using audio terminology, each extra bit of data moves the noise floor down by 6dB (half). We can turn this around and say that each bit of data provides 6dB of dynamic range (*4). Therefore 16bit x 6db = 96dB. This 96dB figure defines the dynamic range of CD. (24bit x 6dB = 144dB).

So, 24bit does add more 'resolution' compared to 16bit but this added resolution doesn't mean higher quality, it just means we can encode a larger dynamic range. This is the misunderstanding made by many. There are no extra magical properties, nothing which the science does not understand or cannot measure. The only difference between 16bit and 24bit is 48dB of dynamic range (8bits x 6dB = 48dB) and nothing else. This is not a question for interpretation or opinion, it is the provable, undisputed logical mathematics which underpins the very existence of digital audio.

On page 272 I can tell this is an interesting subject, I haven't had time to read all of the pages yet so I'm just responding to the original assertion with my thoughts here.

Whenever bit depth is discussed along comes dither as the knight in shining armour, and the noise is always a 'tiny bit'.
For 2 bit audio the noise would in fact be quite significant, but lets leave that aside. In order to include classical music - some of which has a fair bit of music at 48dB down we are not talking about 16bit audio but 8bit audio, so (single bit) dither in that case is 1/255 or 0.4% noise, whereas in a 24bit system 48dB down (24 - 8 = 16) it would be 0.0015%.

But it's not so much even the noise, it's the problem of transient peaks, little transient peaks. Dither is a process for continuous waves because the error gets averaged out over time, this doesn't help transient peaks, which technically dither will damage more than no dither, because you are potentially moving all of the points around by a level (1). That means for a quiet 16bit classical section (which 48db down is 8bit) that 0.4% noise becomes a 0.4 - 0.8% level error for each of the several points of the peak, which I've always considered to be distortion.

This tells us that 24bit will sound better for quiet, transient sounds like a soft cymbal brush, ambient sounds, subtle cues, but for 1kHz test tones there may not be any point to 24bit.
Maybe others are different, but I don't listen to 1kHz test tones, I prefer dynamic music with both loud (fine for 16bit) and quiet (bad for 16bit) parts.

 

[gregorio]

[1] Whenever bit depth is discussed along comes dither as the knight in shining armour, and the noise is always a 'tiny bit'.
[2] For 2 bit audio the noise would in fact be quite significant, but lets leave that aside.
[3] In order to include classical music - some of which has a fair bit of music at 48dB down we are not talking about 16bit audio but 8bit audio, so (single bit) dither in that case is 1/255 or 0.4% noise, whereas in a 24bit system 48dB down (24 - 8 = 16) it would be 0.0015%.
[4] Dither is a process for continuous waves because the error gets averaged out over time, [4a] this doesn't help transient peaks ...
[4a] That means for a quiet 16bit classical section (which 48db down is 8bit) that 0.4% noise becomes a 0.4 - 0.8% level error for each of the several points of the peak, which I've always considered to be distortion.
[5] This tells us that 24bit will sound better for quiet, transient sounds like a soft cymbal brush, ambient sounds, subtle cues ...

1. Dither is not the knight in shining armour, it is intrinsic to digital audio implementation. It's like saying wheels are the knight in shining armour for a car!
2. And for 1 bit audio the noise would be twice as much as for 2 bit. Ever heard a SACD, did you notice that ridiculous amount of noise?
3. Not quite sure how you're working that out or what you're trying to say but yes, I agree, 24bit audio is generally easily distinguishable from 8bit (with music recordings and no noise-shaped dither), due to the amount of audible noise.
4. I think you've misunderstood something at a fundamental level. Dither does NOT "average the error out over time", it effectively converts the error into uncorrelated noise (white noise). The fewer the bits, the more error and therefore more dither is required to eliminate that error and you end up with more white noise but regardless of the amount of dither (and resultant white noise), what we're left with after dithering is ZERO error. So for 16bit, the error is NOT 0.8% it's 0.0%, it's always 0.0% after dithering at any bit depth, which is why dithering is ALWAYS applied and why it's intrinsic to digital audio!
4a. Correct, it doesn't help the transient peaks. However, it doesn't hurt/harm the peaks either. In fact, it doesn't affect the peaks (or any other property/characteristic of the audio) in any way whatsoever, until we get down to the level of the white noise and even then the audio still has zero error, it's just difficult to recognise that fact because it's buried in that white noise. The only issue left then, is how much white noise do we have? For 16bit, that white noise would be at about -92dB, which is more than 100 times below the -48dB figure you're quoting!
5. No, the facts tell us the exact opposite! It tells us that 24bit CANNOT sound better for quiet transients or any other sounds, because whatever bit depth we use the result will always be zero error after dithering! Again, the only issue is how much white noise do we have in addition to our zero error signal but when the level of that white noise is well below the noise floor of the listening environment, the noise floor of the recording chain, the noise floor of the reproduction chain or the noise floor of the recording environment itself, then we can forget about it because it's inaudible and irrelevant. The level of the white noise with 16bit fulfils not just one but all of these conditions!

G

 

[castleofargh]

On page 272 I can tell this is an interesting subject, I haven't had time to read all of the pages yet so I'm just responding to the original assertion with my thoughts here.

Whenever bit depth is discussed along comes dither as the knight in shining armour, and the noise is always a 'tiny bit'.
For 2 bit audio the noise would in fact be quite significant, but lets leave that aside. In order to include classical music - some of which has a fair bit of music at 48dB down we are not talking about 16bit audio but 8bit audio, so (single bit) dither in that case is 1/255 or 0.4% noise, whereas in a 24bit system 48dB down (24 - 8 = 16) it would be 0.0015%.

But it's not so much even the noise, it's the problem of transient peaks, little transient peaks. Dither is a process for continuous waves because the error gets averaged out over time, this doesn't help transient peaks, which technically dither will damage more than no dither, because you are potentially moving all of the points around by a level (1). That means for a quiet 16bit classical section (which 48db down is 8bit) that 0.4% noise becomes a 0.4 - 0.8% level error for each of the several points of the peak, which I've always considered to be distortion.

This tells us that 24bit will sound better for quiet, transient sounds like a soft cymbal brush, ambient sounds, subtle cues, but for 1kHz test tones there may not be any point to 24bit.
Maybe others are different, but I don't listen to 1kHz test tones, I prefer dynamic music with both loud (fine for 16bit) and quiet (bad for 16bit) parts.

the transient things is a mystery to me. I can only agree that 16bit has higher quantization noise than 24bit. we all agree on that.
dither in it's most basic form is there to decorrelate that quantization noise, but noise shaping does in practice the opposite of what you're saying. it provides in principle at least, the ability to perceive signals even below -96db. I know it's not intuitive, dither applied to a picture would probably make more sense than trying to provide an example at such a low audio level, or you could always fool around with a 8bit files and different dither options to try and experience the principle yourself.

now about taking a quiet sound at -48db(assuming full scale would be at normal to loud listening level) within the track and looking at it as if it was somehow the only sound and we have perfect hearing at that level, that doesn't work because in real life situations signal amplitude and quantization noise aren't the only variables involved. for example, what about all the other noises? noises in the studio, noises from the microphones, both most likely increased on some of the tracks to match the gain with the other instruments while mixed. what about your playback system? your listening room? they're all likely to add noises/distortions louder than the dither noise. so you can cry about your low level sounds, because yes they have noises and other stuff creepping just below or even above them. and yes as waves have a cumulative nature, all of it ends up in the final sound wave. but 16bit dither will in practice still be low enough to register most of those noises and play them back to you on a perfect playback system in an anechoic chamber where you might be able to tell how they sound. well you will if they aren't masked psycho acoustically by the -48db music. ^_^

 

[Cutestudio]

the transient things is a mystery to me. I can only agree that 16bit has higher quantization noise than 24bit. we all agree on that.
dither in it's most basic form is there to decorrelate that quantization noise, but noise shaping does in practice the opposite of what you're saying. it provides in principle at least, the ability to perceive signals even below -96db.
I know it's not intuitive, dither applied to a picture would probably make more sense than trying to provide an example at such a low audio level, or you could always fool around with a 8bit files and different dither options to try and experience the principle yourself.

Jiggling a signal that’s not causing a bit transition pushes it over the transition point for an amount statistically proportional to its actual amplitude level.

My argument is that for short transient sounds in music you haven't got enough samples for those statistics to work.
The correct fix for quantisation error in 16bit audio is therefore to switch to 24bit audio, which today is simple and almost without cost.

now about taking a quiet sound at -48db(assuming full scale would be at normal to loud listening level) within the track and looking at it as if it was somehow the only sound and we have perfect hearing at that level, that doesn't work because in real life situations signal amplitude and quantization noise aren't the only variables involved. for example, what about all the other noises?

Classical music has a wide dynamic range with quiet stuff being played at reasonable volume. With a decent amp and speakers: or decent headphones, this is not an unreasonable scenario in the HiFi world.
You appear to be saying 24 bits is more accurate than we need...
...What's the drawback of using a format that's better than we need?

 

[Cutestudio]

Dither does NOT "average the error out over time", it effectively converts the error into uncorrelated noise (white noise). The fewer the bits, the more error and therefore more dither is required to eliminate that error and you end up with more white noise but regardless of the amount of dither (and resultant white noise), what we're left with after dithering is ZERO error. So for 16bit, the error is NOT 0.8% it's 0.0%, it's always 0.0% after dithering at any bit depth, which is why dithering is ALWAYS applied and why it's intrinsic to digital audio!

I understand how this applies to continuous tones, but not how it applies to transients.
That "uncorrelated noise" is the error being spread (and resolved) over time - is it not?
If so, don't we need the error fixing at the time of the transient, rather then over the next few hundred samples?

 

[bigshot]

What kind of transient are you going to find in music that doesn't cover hundreds of samples? I'm just guessing, but I would bet a very fast snare drum hit probably lasts no less than 20 ms from attack to decay. Even if you just worry about the attack, that's still 2 or 3 ms.

 

[pinnahertz]

What kind of transient are you going to find in music that doesn't cover hundreds of samples? I'm just guessing, but I would bet a very fast snare drum hit probably lasts no less than 20 ms from attack to decay. Even if you just worry about the attack, that's still 2 or 3 ms.

You're of by at least an order of magnitude. What frequency would have a 20ms period? That's 50Hz. The attack portion of a snare hit is much faster, the attack of a cymbal, bell or glockenspiel is up higher than 20kHz (the decays are of course much longer). The fastest single transient that can be recorded accurately would have a period of 1/2 the sampling frequency, for 44.1 that's about 45uS, and that would be two samples. If a transient with that kind of attack decays a while, it becomes many samples quickly.