Testing audiophile claims and myths

[bigshot]

It really doesn’t matter because even without dither, you’re not likely to hear the noise when you’re listening to music. There’s a test of that in the Ethan Winer link in my sig.

 

[71 dB]

It really doesn’t matter because even without dither, you’re not likely to hear the noise when you’re listening to music. There’s a test of that in the Ethan Winer link in my sig.

Yeah, dither is more of a "doing things the right/smarter way" thing rather than being something absolutely necessary.

If we have a 24 bit file, but it has a noise floor at say -90 dBFS, this noise will act as self-dither when truncating to 16 bit and additional dithering isn't needed.

 

[gregorio]

2. The modified signal cannot be more accurate than the original unmodified signal statistically is my point.

Then your point disagrees with the well established/proven facts! To name only two of the supporting papers, the first of which was presented in 1984:
“Dither in Digital Audio” - Vanderkooy, Lipshitz.
“The Theory of Dithered Quantization” - R Wanamaker.

error is cancelled out, sure. But now it is on integer 7 on the 5th sample of the dithered signal. It should be integer 8.

After rounding:
original 88888
dithered 88887

This is what doesn’t make sense. What original “88888”? There would never be an 88888, dither has always been applied to the input signal, no commercial ADC exists or has ever existed that does not apply dither, so the output would always be 88887, there would never be an 88888 “original”. 88888 is a purely theoretical output, that would only exist if the input signal were not dithered and were rounded instead. You seem to think there is some 88888 integer sequence somewhere in the process that is then modified with dither, this is not the case, it never exists.

What doesn’t make sense is your assertion that “it should be integer 8”, when you stated the actual value is 7.8. Even just beginner math indicates that it should not be integer 88888, it should be 88887. 88888 represents a statistical error, 88887 does not.

Statistically there is trading some integer accuracy there.

What integer accuracy? By your own admission, 88888 would NOT be accurate, it would be inaccurate, 88887 is accurate. What you’re “trading” is therefore some integer inaccuracy (with 88888) that you would get if you rounded or truncated instead of dithered, for perfect accuracy (88887). In terms of analogue output signal result, after the DA Conversion process, you are trading deterministic/periodic correlated distortion for uncorrelated noise (which is inaudible).

There's no way around it.

Why on earth would you want a “way around” getting a sequence of accurate integers (instead of inaccurate integers)? What we want is the exact opposite, a “way around” getting inaccurate integers, which is exactly what dithering gives us!

G

 

[danadam]

it's a trade off. less noise for less accuracy.

More like: less distortion for more uncorrelated noise (white or shaped). As for accuracy, when you convert signal at level below least significant bit, without dither you get zero, with dither you get noisy signal. Are you saying that "no signal at all" is more accurate than "the signal" with uncorrelated noise?

 

[eq1849]

Then your point disagrees with the well established/proven facts! To name only two of the supporting papers, the first of which was presented in 1984:
“Dither in Digital Audio” - Vanderkooy, Lipshitz.
“The Theory of Dithered Quantization” - R Wanamaker.

This is what doesn’t make sense. What original “88888”? There would never be an 88888, dither has always been applied to the input signal, no commercial ADC exists or has ever existed that does not apply dither, so the output would always be 88887, there would never be an 88888 “original”. 88888 is a purely theoretical output, that would only exist if the input signal were not dithered and were rounded instead. You seem to think there is some 88888 integer sequence somewhere in the process that is then modified with dither, this is not the case, it never exists.

What doesn’t make sense is your assertion that “it should be integer 8”, when you stated the actual value is 7.8. Even just beginner math indicates that it should not be integer 88888, it should be 88887. 88888 represents a statistical error, 88887 does not.

What integer accuracy? By your own admission, 88888 would NOT be accurate, it would be inaccurate, 88887 is accurate.

G

I'm talking 24bit to 16 bit down conversions. If that wasn't clear, my apologies.

So to be clear:

We take a 24 bit file and down convert it to a 16 bit file and the signal(to keep it easy it's a constant signal of 7.8) over 5 sample points is:
7.8, 7.8, 7.8, 7.8, 7.8
8,8,8,8,8...after rounding to the nearest integer
rounding error of 1=(8-7.8)*5

If we apply dither it rounds to(As per 71 db's example dither rounds down 20% of the time rather to the nearest integer)
8,8,8,8,7
rounding error of 0=((8-7.8)*4+(-.8))


Conceptually, is that correct? Correct me if something is off. I'm happy to learn.

 

[bigshot]

Origin of Big Endian and Little Endian
https://www.ling.upenn.edu/courses/Spring_2003/ling538/Lecnotes/ADfn1.htm

 

[gregorio]

I'm talking 24bit to 16 bit down conversions. If that wasn't clear, my apologies.

Ah OK, that’s slightly different.

Conceptually, is that correct? Correct me if something is off. I'm happy to learn.

No it’s not correct but your example “to keep it easy” oversimplifies to the point that it doesn’t represent what actually happens, it doesn’t really make sense and it’s difficult to explain in it’s own terms. The conversion from 24bit to 16bit without dither is simply a truncation of the 8 least significant bits (LSBs). So if our actual value were 7.8 in 24bit and our decimal point represents the point of truncation (after the 16th bit), then the result of conversion without dither would be 77777, not 88888, there’s no rounding.

In the real world we’re dealing with thousands/many millions of samples rather than just 5 and there will only ever be a sequence of identical values (integers) in two cases: digital silence or continuous overload clipping (all bits set to one). Discussing everything apart from these two conditions and trying to use your example, the first sample in 24bit will be say 7.8000000, the next will be say 7.7543219, the next say 7.6243153, etc. The standard way to reduce the bit depth to 16bit would be with noise-shaped dither, which would result in integers (within the critical hearing band) that statistically will result in values identical to say 7.80, 7.75 and 7.62, etc. In theory, we could in fact get back our identical 24bit values (to all decimal places) but that’s impractical with a 44.1kHz sampling freq. Regardless, that’s obviously a great deal more accurate than just the 7, 7, 7 we would get from truncation.

In short, your conceptualisation is incorrect, not only because it’s a gross oversimplification but also because it’s internally also incorrect (there’s no rounding for example).

This link is about as much of a simplification as is reasonable, in order not to throughly misunderstand the issue. Note that it’s also interesting from the point that it’s a highly respected manufacturer who’s giving precise details of what their equipment is actually doing, which is a rarity and lastly, it represents the “state of play” around 20 years ago: “Dither and Noise-Shaping” - Prism Sound, c. 2006.

G

 

[71 dB]

If we assume 16 bit being integers (0000 0000 0000 00012 = 110), then the last 8 bits in 24 bit represent the stuff on the right side of the desimal point. Interestingly 7.8 needs infinitely many bits to be 100 % accurately represented in binary:

111.110011002 = 7.79687510
111.110011012 = 7.8007812510 <= closest to 7.8

7.7510 is much easier in binary: 111.112.

 

[eq1849]

Ah OK, that’s slightly different.

No it’s not correct but your example “to keep it easy” oversimplifies to the point that it doesn’t represent what actually happens, it doesn’t really make sense and it’s difficult to explain in it’s own terms. The conversion from 24bit to 16bit without dither is simply a truncation of the 8 least significant bits (LSBs). So if our actual value were 7.8 in 24bit and our decimal point represents the point of truncation (after the 16th bit), then the result of conversion without dither would be 77777, not 88888, there’s no rounding.

In the real world we’re dealing with thousands/many millions of samples rather than just 5 and there will only ever be a sequence of identical values (integers) in two cases: digital silence or continuous overload clipping (all bits set to one). Discussing everything apart from these two conditions and trying to use your example, the first sample in 24bit will be say 7.8000000, the next will be say 7.7543219, the next say 7.6243153, etc. The standard way to reduce the bit depth to 16bit would be with noise-shaped dither, which would result in integers (within the critical hearing band) that statistically will result in values identical to say 7.80, 7.75 and 7.62, etc. In theory, we could in fact get back our identical 24bit values (to all decimal places) but that’s impractical with a 44.1kHz sampling freq. Regardless, that’s obviously a great deal more accurate than just the 7, 7, 7 we would get from truncation.

In short, your conceptualisation is incorrect, not only because it’s a gross oversimplification but also because it’s internally also incorrect (there’s no rounding for example).

This link is about as much of a simplification as is reasonable, in order not to throughly misunderstand the issue. Note that it’s also interesting from the point that it’s a highly respected manufacturer who’s giving precise details of what their equipment is actually doing, which is a rarity and lastly, it represents the “state of play” around 20 years ago: “Dither and Noise-Shaping” - Prism Sound, c. 2006.

G

Are dither and noise shaping one and the same?

 

[gregorio]

Are dither and noise shaping one and the same?

Not entirely, “noise-shaping” is the process of shaping (redistributing) the noise produced by dither. There is no noise-shaping without dither, although there is dither without noise-shaping. When we say “noise-shaping”, we mean “noise-shaped dither”.

G

 

[Ruben123]

Not really on-topic but since this is a sort of science cafe.... I just read a topic (gregorio against some other dude in 2020) where it was discussed how much a $10,000 Ethernet cable improved sound and that AB/X'ing it was regarded as "too much bother". I really am out of words.

 

[gregorio]

I really am out of words.

Indeed, it often is difficult to think of the words. You’ve either got to be so wealthy that thousands of dollars really is just a drop in the ocean or exceptionally ignorant and deluded. The problem in the audiophile world is that level of ignorance and delusion isn’t exceptional, it’s actually so common it’s “normal” and therefore anyone who disagrees must be wrong/“abnormal”.

G

 

[Joe Bloggs]

Not really on-topic but since this is a sort of science cafe.... I just read a topic (gregorio against some other dude in 2020) where it was discussed how much a $10,000 Ethernet cable improved sound and that AB/X'ing it was regarded as "too much bother". I really am out of words.

I would dismiss any difference I hear as absolute placebo and regard ABXing it as too much bother, but that's just me.

 

[gregorio]

I would dismiss any difference I hear as absolute placebo and regard ABXing it as too much bother, but that's just me.

And most others who aren’t “exceptionally ignorant and deluded”. Most with a reasonable understanding of Ethernet wouldn’t bother either, as how it works essentially precludes the effects claimed by some audiophiles.

G

 

[mbphotox]

same here..
I'm too lazy to bother ABX-testing most things..

If I end up enjoying something, that's enough