[tdog]

Using dither is best practice. It's virtually free these days and there is no penalty for using it. You can choose not to use it and get inferior results. Many recordings are pretty poorly made, that is no reason to advocate poor practice just because most people won't notice.

 

[bigshot]

It's easy enough to do. But I really don't think it makes a whole heck of a lot of difference. (see Ethan Winer's tests in my link below)

 

[tdog]

Well, it makes a difference _sometimes_, and that's the point. I've chatted with both Monty and Ethan over email about this (and other issues) just last year and they did agree that it's something that is so good and free that there's no reason not to do it. 16-bit resolution is indeed excellent, better than most people need for most compressed music, but for no cost you can have better within the same bit-depth (i.e. 24-bit dithered to 16-bit) and it makes a difference sometimes and to some people, so why not?

 

[limpidglitch]

Besides, Greenears wanted to give 24bit every reasonable favour, so dither goes out the window, precisely because it's 'best practice'.

 

[RRod]

  Well, it makes a difference _sometimes_, and that's the point. I've chatted with both Monty and Ethan over email about this (and other issues) just last year and they did agree that it's something that is so good and free that there's no reason not to do it. 16-bit resolution is indeed excellent, better than most people need for most compressed music, but for no cost you can have better within the same bit-depth (i.e. 24-bit dithered to 16-bit) and it makes a difference sometimes and to some people, so why not?

 
I don't think anyone disagrees.

 

[Don Hills]

... Yes I've read most CD are mastered with peaks at -10 to -15 dBFS.   So let's say this one peaks at -10 dBFS and with 50dB DR then the bottoms are at -60dBFS and let's say the noise is at -90.  You only have 30 dB SINAD for the soft parts compared to loud parts 80 dB.  It's exactly the same thing I just said before, expressed in dB.  Since I expect 80 dB is not audible, I focus on the 30.

 
You must have read a very old book. Very few CDs are mastered with peaks that low nowadays. Certainly no popular ones.
Also, when you are doing your listening tests, remember that you aren't supposed to turn up the volume for the quiet parts. You set the volume to your normal listening level for the genre of music you're listening to and leave it at that.

 

[macshooter]

I'm don't have time to read all 134 pages of this thread to see if someone else has already tossed this out there (I would be surprised if no one had) but this article really opened my eyes a bit on this topic. Some reviews of some interesting audio equipment on that site as well.

 

[stv014]

 
Using dither is best practice. It's virtually free these days and there is no penalty for using it.

 
Actually, there is some penalty for using dither, since it increases the noise floor, and with real music (that is, not a single pure tone, or something so quiet that the quantized version will only have information in a small number of bits) that does not have a very high dynamic range, the undithered quantization error will be mostly noise anyway. See also these older posts. With 16-bit samples, it usually just does not matter either way, because neither the quantization nor the dither noise is audible under normal conditions. The perceived noise level can be reduced with noise shaping, although at low sample rates like 22050 Hz it can do more harm than good.
 
Dithering may be best practice, as it allows for a more consistent and "analog-like" performance (i.e. a guaranteed uncorrelated noise floor), but do not assume that an undithered sample is necessarily "broken", or even sound worse overall (other than e.g. when listening a very quiet section like a long fade-out to silence at the end at a volume setting that would be unbearable for the rest of the track) than one with simple non-shaped dither.
 

 
Truncating 24-bit audio to 16-bit is something that no one should do these days, even for relatively compressed and low dynamic range pop music - the difference is audible on good playback systems, as per the links in my previous post.

 
I doubt you would be able to hear the difference with compressed pop music without some form of cheating (which includes any use of volume settings unsuitable for listening to the entire track, and/or pre-attenuating the sample before quantizing it so that the signal will have very little entropy left). In the well known Meyer and Moran tests, no dithering was used, and, while there was some criticism regarding the samples chosen, they were generally better than compressed pop music.

 

[stv014]

  Besides, Greenears wanted to give 24bit every reasonable favour, so dither goes out the window, precisely because it's 'best practice'.

 
In this case it might not matter that much, though, since ABX only tests transparency, rather than which sample sounds "better". In cases where non-dithered quantization is not transparent, the noise floor of non-shaped dither has good chances of being audible as well, and therefore lead to a positive ABX result, even if it is aesthetically less objectionable than distortion.
 
The main problems seem to be statistical bias (doing multiple runs until one produces a p-value <= 0.05, which will sooner or later happen by simple luck), and listening to quiet sections at unrealistically high volume.

 

[limpidglitch]

   
In this case it might not matter that much, though, since ABX only tests transparency, rather than which sample sounds "better". In cases where non-dithered quantization is not transparent, the noise floor of non-shaped dither has good chances of being audible as well, and therefore lead to a positive ABX result, even if it is aesthetically less objectionable than distortion.
 
The main problems seem to be statistical bias (doing multiple runs until one produces a p-value <= 0.05, which will sooner or later happen by simple luck), and listening to quiet sections at unrealistically high volume.

 
In fact we expect him to manage that in 1 out of 20 tries. Statistics is weird like that 

 

[Greenears]

Thanks all for the interesting SoX incantations. I really appreciate it.
 
I'm certainly not the OP (134 pages into this thread!) but I'm the OP on this mini sub-thread I guess.  Let me kindly restate my purpose:
 
I am not a fan of ABX tests where someone does a series and gets fail-fail-fail.  I have seen many many such logs online and they leave me cold.  This is just my perspective.  Because it begs the question whether the tester has good enough ears, good enough transducers etc.  IMO you have to pass something first, and then gradually dial up the difficulty until you fail.
 
So my purpose here is to do something relatively pure - you'll notice I converted the 24 bit myself rather than relying on the HD tracks 16 bit version, since you don't know how that was done.  But after that, give 24 bit every chance to win.  Look at the results as you go.  Crank the volume on soft passages.  Listen to a very short clip over and over. Everything other than looking at the electronic signature.  I know you consider this "cheating" but it is not.  If your ears can detect it, fair game.  I'm not trying to reproduce "natural listening" in my first pass.  I promise once I get a pass I'll start to dial back on these things.  Then people who view the logs can make their own decision whether it matters to them.  
 
Please realize I fully understand the mathematics behind dither, normalization, sampling, FIR filters, and all kinds of other semi-useless information.    None of your comments are invalid, I'm just saying not yet. I am still struggling to see the magic 5% first.
 
Which me luck! I hope to try again later today.  I also set up and converted some Linn tracks yesterday.  I encourage anyone interested to try to reproduce.   All the tracks and software I'm using is 100% free.  If anyone has a different passage or track that passes for them please let me know.

 

[bigshot]

  IMO you have to pass something first, and then gradually dial up the difficulty until you fail.

 
I'm learning to fly like that. I started by standing on a chair and flying to the ground. I'm working my way up to jumping off the top of the Empire State Building.

 

[Don Hills]

   
I'm learning to fly like that. I started by standing on a chair and flying to the ground. I'm working my way up to jumping off the top of the Empire State Building.

He does have a point though. The test should have positive and negative controls. Samples where the test subject should be able to tell a difference, and samples where he should not.

 

[Greenears]

   
I'm learning to fly like that. I started by standing on a chair and flying to the ground. I'm working my way up to jumping off the top of the Empire State Building.

Oh how humorous we are .... but thanks for proving my point.  If you practice flying by increasing 3 feet at a time, at least we learn how high you can fly.  With ABX testing all I see is a bunch of people jumping from different skyscrapers - all with the same predictable result.  That is why you don't learn that much, except what you already expected. 
 
Remember the only thing you can absolutely learn from ABX testing is that if you pass you were absolutely NOT guessing (down to extremely unlikely probability).  A fail tells you close to nothing - A could have been Bach and B Mozart and maybe the listener was just not paying attention.

 

[RRod]

  Oh how humorous we are .... but thanks for proving my point.  If you practice flying by increasing 3 feet at a time, at least we learn how high you can fly.  With ABX testing all I see is a bunch of people jumping from different skyscrapers - all with the same predictable result.  That is why you don't learn that much, except what you already expected. 
 
Remember the only thing you can absolutely learn from ABX testing is that if you pass you were absolutely NOT guessing (down to extremely unlikely probability).  A fail tells you close to nothing - A could have been Bach and B Mozart and maybe the listener was just not paying attention.

 
It all comes down to the listeners intentions. If I want to fail an ABX deliberately, I just pick A every time. But if someone said "I will pay you $100 if you can pass", then I'm listening intently to find the difference. It's fine to futz around with improper variations of a test to understand its inner workings, but eventually you have to actually take it in earnest.