PDA

View Full Version : Steve Massey says "don't bother with dither"?!



Tarktones
05-02-2008, 03:57 PM
I was recently reading a cookbook. The author was emphasizing the importance of using good salt over cheaper, generic salt. He described the flavor differences and insisted it had an important impact on the final dish. I chuckled with skepticism. But, I happened to have two different brands of salt in the cupboard. One was a domestic brand, a sea salt of an unknown origin. Its container described how it was better because it didn't have chemical bleaching agents and retained the natural trace minerals. The other was a more expensive imported naturale Italian sea salt. So, of course, I performed a taste test. To my huge surprise, there was a very significant, undeniable difference in taste. The locations of sensation on the tongue were also different. The Italian brand easily won out. It was smooth, pleasant, and the flavor finished evenly. By comparison, the other brand was strident, piercing, and had a synthetic chemical-like flavor and feeling. Both immediately evoked the idea of salt -- but, side-by-side, the differences were almost like that between real sugar and an artificial sweetener. I presented this taste test to a handful of other people, and all the reactions have concurred with mine. I will certainly never think, or rather not think, of salt the same. The cookbook writer was correct. I've found it does make a subtle but worthwhile impact on the end product, making the eating experience just a little bit more enjoyable, especially after all that hard work in the kitchen.

So what does salt have to do with audio dithering? Dither and salt both speak to the issues of marketing, personal enjoyment, and conventional wisdom.

Conventional wisdom says that all salt tastes the same. It's salt! NaCl. Sodium chloride. How could one box of NaCl taste that much different from another box of NaCl? Conventional wisdom scoffs at the idea that one brand might taste much better than another. Conventional wisdom laughs at the idea of spending five dollars for a can of it, rather than a buck for that old Morton's.

Conventional audio engineering wisdom says dither's a bit like salt -- you add a pinch of noise to "enhance the flavor." Conventional wisdom says that you'd be crazy not to dither when mixing down to a 16-bit master. I say: Conventional wisdom isn't all that wise. Stop bothering with dither! Unlike salt -- it has no impact on the end product.

I know what a lot of people will say to this: But, I've heard the improvements of dither with my own ears! I've done the taste test. I've read the book. And, I've seen the movie!

For example, you can go to this website and evaluate all the major players in dithering technology. It's true -- you can absolutely hear the improvements in their examples. And, that particular algorithm with the most votes does sound the best. But, let's read the fine print about the chemical bleaching agents: "54 dB of gain has been applied after dithering." Since the audio isn't audibly clipping, this means there was at least 54 dB of headroom in the original digital signal. 54dB of gain equals a multiplication factor of 501. In binary, this equates to shifting the digital word almost 9 bits (501X = 2 to the power of 8.97 bits). In other words, the original audio had about 9 unused bits of headroom. 16 minus 9 equals 7. We're listening to the equivalent of 7-bit audio here. So, what does this listening test prove?

That dither is highly effective for 7 or 8 bit audio. How many recording engineers do you know sending out 8 bit masters?

Why did the creators of this test use such bit-reduced audio? They had to or the test never would have worked. Quantization noise, which dither is meant to mask, is not humanly audible for typical, normalized, 16-bit audio, especially not in the everyday listening environments where we experience music. 16-bit audio has over 90dB of dynamic range -- this is the difference between the maximum signal amplitude and the quantization noise. In other words, the quantization noise is 90dB below what you're actually listening to. That's huge! Think about that in real world terms: 85dB SPL is a standard peak calibration level for movies and is also the threshold of hearing damage, at long durations. Under this condition, 90dB below that is well under the generally accepted "threshold of human hearing". Expecting to hear the effect of dither is like expecting to hear the fluttering of a fly through the din of a jackhammer -- it's just not going to happen. Maybe... just maybe, with a highly un-normalized Classical CD from the 80's listened to in an anechoic chamber with the lowest-noise-possible audio reproduction gear, you just might barely detect hints of quantization noise on the quietest passages. But, I'm dubious -- who do you know that has an anechoic chamber for a living room anyway?

To further this assertion, why is it so difficult for audio engineers to figure out how to apply dither properly? There's been numerous magazine articles, white papers, tidbits in the many home recording books, and thread after thread on the internet forums asking: "How do I dither?", "When do I need to dither?" , "Should I re-dither?", "Does dither work with floating-point audio, or only fixed-point?", "On what insert do I put the dither plugin?" "Can I EQ the audio after dithering?" And, these questions never end.

If dither makes such a perceptible improvement to fidelity, then shouldn't it be completely obvious when that plug-in gets latched into the correct slot and the audio is flowing through the correct path? Shouldn't the standard, somewhat flippant, internet forum answer of "Just use your ears!" be applicable here? It's not -- no one ever says this with regards to dithering. That's because, it's hard to stand behind such a statement with any confidence about a technology that does not exist.

A while back I discovered that the TDM POW-r dithering plugin that's provided with Pro Tools had a bug, making it incompatible with my L2007 limiter plugin. The POW-r plug-in would stop producing dither when activated on the same DSP as the L2007. (My limiter, unlike most others, does not provide integrated dithering, so I've always suggested to people that they just use the POW-r plugin.) I discovered the bug on my own, more or less by accident, while using an FFT analysis plugin. But, guess how many users called to report it themselves? Zero. Just like myself, no one ever heard this on their own. Only with tweaky analysis tools is dither measurable.

Finally, let's consider the end product. When have you ever listened to a CD and thought, "Wow! This is a really incredible album. But, damn! The dithering algorithm that the mastering engineer used is total crap! I can't believe they used POW-r type 3 instead of type 2!" Never. No one has ever had this thought. Regardless of its questionable efficacy, no one can reasonably argue that dither has an impact on the experiential enjoyment of music. Unlike other aesthetic choices of audio production (compression, distortion, etc.), no one has ever argued that dither ruined or enhanced their ability to enjoy a musical work.

The casual listener is capable of perceiving all subtle aspects of music production, whether or not they speak the language of audio engineering. It may be difficult for them communicate what they are hearing. If the vocal is slightly distorted, or if the phase is mutated due to excessive equalization, they may wave their hands around or find a peculiar phrase that best fits the tenuous categorization scheme their mind has created for this language-less reality. Likewise, we can communicate and agree about the flavors of salt, even though no one has previously defined a rigorous categorization scheme of salt. But, dither -- the only people on the planet perceiving and discussing dither are those people who have been informed of its existence -- because dither does not exist in anyone's reality until someone gives it a name.

So, where does the mythology of dither originate? Well, the problem is, it is a mathematical truth. There's no denying that. You can write down and solve the equations. Dither does work in the abstract sense: it decorrelates the quantization noise from the carrier signal. But, again, this theory is only applicable to the human experience for very low-bit, highly-quantized audio signals -- simply put, when the quantization noise gets "really loud", you can hear it.

If dithering makes no difference, how has the concept lived on for so long? I think because it is mathematically validated, it has allowed the engineers at pro audio companies to say, with confidence, to their marketing departments that the inclusion of dithering has made their product better over the competition. That's great! I imagine marketing people love positive specifications and catch-words which they don't have to create themselves. Anything that is easily quantified and can be succinctly composed into the marketing text is great for the unimaginative salesperson. Examples of such marketing-driven myths are common in the technology market. 64-bit floating-point audio! Awesome! How in the world could that not be better than 32-bit floating-point audio? 64 is twice the size of 32. Six Megapixels is of course better 5 megapixels! (Like with most bullet-points and talking points, these metrics ignore the subtleties and/or disregard all the other equally important elements that were compromised in reaching that isolated specification.)

But, I think the primary reason dither lives on is because the public itself has embraced it wholeheartedly, beyond simple influences of marketing. Why? I believe it's a bit like humanity's attachment to the supernatural. People seem to posses, myself included, an overwhelming desire to imagine their perceptions of the world to be more subtle and more magical than they actually are. Even though no one has ever truly witnessed the effects of dither, the public has faith in its power. And, since it's impossible to prove or refute something that you cannot hear (or see, or touch), the ghost of dither lives on...

Interesting perspective.

DAS
05-02-2008, 04:21 PM
There is some validity to this, but there are also a number of places where it completely falls apart and is wrong. As is often the case, it depends.... One thing for sure, a lot of people do get overly worked up over dither on material that will be played back in an environment where it couldn't possibly make one iota of difference. Unfortunately this is written in a way that leaves room for people to generalize about it ("dither is bad," or at least "not necessary"), which is not correct in many circumstances. I don't have time to elaborate right now, but I wouldn't suggest you turn off all those dithers quite yet. But feel free to experiment. It's actually not that hard to do some tests in your average listening environment.

Rad
05-02-2008, 04:46 PM
I fully agree with DAS.

This article uses facts that aren't invalid in themselves, but ultimately twists them in order to make a political point which is completely unjustifiable.

In particular, the claim that the differences in dither are audible only in situations with huge gain (as in the example he provides) is, of course, completely wrong. This just isn't true.

Many of the Waves dithers, for example, sound different even on regular listening volumes, not to mention that the new L3 now offers not only different dithers but also different noise-shaping algorithms (Loud and Proud, Analog, Warm, etc) which make a very audible difference on how reverb trails and quiet passages sound.

If you're listening on a boombox, granted, you won't hear the differences.
If you're an audiophile listening at home, most likely you will.
Why should engineers shoot just at the average (or even below average) audience and how many people today listen on boomboxes?

aitikin
05-02-2008, 05:25 PM
...and how many people today listen on boomboxes?

Not many, but all too many listen on iPod headphones...

Rad
05-02-2008, 06:35 PM
Sure. These folks most likely will not hear a difference.

Yet if we start targeting the iPod quality though, then there's no need to use 16 bits either. Twelve would be plenty.

What I mean is that the article is correct that for many everyday users it doesn't make a difference, but where it gets it seriously wrong is the argument that the target quality for the studio engineer should be that of the average user.

You can always take a high-quality audio and downgrade it to, say, an mp3, but the opposite is impossible. That is why I am disagreeing with Massey. And in fact, I know a lot of industry professionals which would disagree too.

Dave Burris
05-02-2008, 06:41 PM
Dither is primarily to mask quantization errors for low-level signals and should generally only be done during the final mix. In many cases the signal chain is inadequate to really benefit that much. You will mostly notice it on higher-frequency material during low passages where the quantization noise stands out most.

Tarktones
05-02-2008, 07:53 PM
I can honestly say I've never heard quantization noise; distortion yes but noise, no. I've never spent much time listening for it either.

Rad
05-02-2008, 08:14 PM
On a 16-bit system, quantization noise is below any audible levels. It is "stored" in the last bit, which means that it lives between the -90-th and the -96-th decibel.
The point of dither is not to fight quantization noise as such and it will not make it either less or more audible. It is to mask the distortion present in the quantized signal, which is sometimes audible. The human ear notices distortion much more easily than noise.

dpd
05-04-2008, 12:05 AM
well, I wonder if Massey dithers in between processing steps in his plugins or simply rounds off between them...

EC_Beast
05-04-2008, 10:41 AM
How does Steve then, intend on getting a 24bit record, down to 16? It wont just get there by itself. Does he suggest some other converter, just recording in 16bit?

Rad
05-04-2008, 05:22 PM
I figure what he's suggesting is truncation. You take the 24-bit word and you just read the first 16 bits. This is just guaranteed to generate some distortion.
I'm amazed this can be taken as "a matter of personal taste" as he seems to be doing with all these salt comparisons. For example, to quote once again from one of my favorites on digital audio, Nika Aldrich, since he says it clearly:

"The audible sound quality degradation of many of the early digital mixers can be traced to the lack of dither."

Clear and concise! Plus dispells the myth that quantization distortion is not audible.
Maybe if you skip the final dithering step, it isn't very audible, altho it's there. But if you've got a chain of 7-8 plugins each of which processes internally at 32 bits but does no dither, just sit and watch how your sound quality will degrade.

dpd
05-04-2008, 10:54 PM
I can't imagine Massey suggesting truncation. That's just stupid. Rounding in place of dither, but not truncation.

TimOBrien
05-05-2008, 08:50 AM
Simple solution: Try it both ways; use whatever sounds best to you.

There, that wasn't so hard was it????

(Too many people just blindly follow what they're told without trying it out themselves!!)

DAS
05-05-2008, 09:42 AM
Dither is primarily to mask quantization errors for low-level signals and should generally only be done during the final mix.

This is also a generalization that breaks down. The one generalization that holds up is: Dither of the proper type and amplitude should technically be applied anytime bit depth reduction takes place. It may not always make an audible improvement in the final result, but under the right circumstances it can.

Rad
05-05-2008, 10:45 AM
I can't imagine Massey suggesting truncation. That's just stupid. Rounding in place of dither, but not truncation.

From the original article, at least for me it's hard to tell what exactly he has in mind. The one thing I agree with is that obsession with dither is harmful. Otherwise, rounding may work. Rounded stuff will still be correlated with quantization error, but I think likely this is not very audible.

Dave Burris
05-05-2008, 04:56 PM
This is also a generalization that breaks down. The one generalization that holds up is: Dither of the proper type and amplitude should technically be applied anytime bit depth reduction takes place. It may not always make an audible improvement in the final result, but under the right circumstances it can.

True. However, in Cubase (for example) the mix engine and the VST effects use 32-bit floating point that maintains 24 significant bits and a mantissa. Seems a stretch to me to interpret that as changing bit depth if you are storing back to 24-bit integers unless you are doing funky things with the gain stages in the signal chain.

In the end there's no real substitute for knowing what's happening in your signal chain, and yes dithering should occur when there is a reduction in the bit depth. Personally, I do not consider the conversion between 24-bit integer and 32-bit float to be in that category.

DAS
05-05-2008, 05:05 PM
True. However, in Cubase (for example) the mix engine and the VST effects use 32-bit floating point that maintains 24 significant bits and a mantissa. Seems a stretch to me to interpret that as changing bit depth if you are storing back to 24-bit integers unless you are doing funky things with the gain stages in the signal chain.

In the end there's no real substitute for knowing what's happening in your signal chain, and yes dithering should occur when there is a reduction in the bit depth. Personally, I do not consider the conversion between 24-bit integer and 32-bit float to be in that category.

Correct, not only is that not a true conversion of bit depth and thus does not require dither, it's actually impossible to effectively apply dither to floating point signals anyway. Think about when a signal leaves an accumulator (at maybe 60 bits or more) from processing a plug-in to return to the 24 bits of precision in the 32-bit floating mixer. Normally you'd think dither should be applied here, and it should, but there just isn't a fully effective way to dither floating point data. This presents a very real problem in 32-bit floating systems, although I would argue that problem exists below the practical noise floor 95% of the time.

Rad
05-05-2008, 06:01 PM
Finally someone who can tell me what exactly Cubase does on the inside! Thank you people! I've wondered multiple times how they do the bit reduction if the internal processing is at more than 24 bits and have never been able to get any definitive answer.