Moulton Laboratories
the art and science of sound
Taking Stock: How Important Is High Resolution Audio, Anyway?
Dave Moulton
September 2000

An epic discussion of high resolution digital audio.

1 2 3 4 5 6 >

Taking Stock: How Important Is High Resolution Audio, Anyway?

Our Story To Date

For the past year, I’ve been ranting about the so-called audibility of our brave new high resolution audio formats, specifically, 24-bit digital resolution and 96 kHz. sampling rate (48 kHz. bandwidth).

I’ve questioned the importance of these “advances” in terms of their audible impact, and I’ve shared with you some of my experience as a tester of “subjective” quantities such as audibility. We’ve discussed, for instance, the laboratory definition of “threshold of audibility” (something that some of us can hear, some of us can’t), the need for blind testing (we’re a sneaky and highly prejudiced lot, by nature, and we can’t be trusted), the errors such blind testing introduces (it’s harder to hear problems when you don’t know they’re there, and that’s unfair), the lunatic hype that emerges from us Golden-Eared types when we finally DO hear something (“Why, I Can Hear It! Wow! It’s Totally Awesome!!”), and a little about the difficulties of making an objective determination of audibility that is in fact “distinguishable from chance” for the kind of small audio differences we’re talking about (takes a heap o’ trials, lots more than most of us really want to do, any given Saturday).

We also took a brief look at our implementation of 24-bit audio, and noted that there is a serious mismatch between the correlation of levels in the digital, analog and acoustic worlds, a mismatch that relegates all that expensively acquired digital resolution to the vanishingly small end of the analog resolution range, as well as buried deep in the perceptual murk that lies beneath the so-called “threshold of audibility.”

As if that weren’t enough, we also looked at the bandwidth question, and noted that there is virtually no evidence that frequencies above 20 kHz. are in fact audible for humans (although rumors abound). Further, we observed that few microphones or loudspeakers in common use have significant bandwidth above 20 kHz., and that there is a fundamental tradeoff in microphone design between noise and bandwidth that effectively limits our efforts to extend both.

Finally, we considered the widely held belief that increased resolution leads to improved accuracy, and that improved accuracy leads to more realistic recordings. We found that current recording/playback practices hardly support accuracy at all in absolute terms, and only slightly in relative terms. I concluded that, so far as I am concerned, “strength of illusion” is a much more important goal in recording than “accuracy.”

What Does It All Mean?

So now that I’ve managed to offend just about everybody, let’s try to put this in a little perspective.

Digital audio, in comparison to analog audio, isn’t a separate parallel system. Rather, it is a subset that exists within analog audio. All audio signals that are digital came from the analog realm (except for a few digital test signals), and in order to listen to them we are going to need to transduce them back into analog form prior to transducing them into the acoustical realm.

The primary points where we have problems with accuracy, resolution, errors and so forth are the points where transducers are used. Generally speaking, A/D and D/A converters (which are, after all, transducers) work pretty well, especially in comparison to microphones, loudspeakers, styli and tape heads. Nonetheless, converters do introduce errors. So, by definition, digital audio can never be better than its analog source, and in effect it will be worse by the amount that those converters add errors. Hence, it behooves us to make those converters work so well that the errors they add are small enough to be, well, “inaudible.”

This is a laudable goal, and in fact it is this goal that has led to those 24-bit 96 kHz. converters and digital audio systems. Why? Because we have come to the conclusion that the errors we encounter with 16-bit 44.1 kHz. systems are in fact “audible.”

And this is where the ozone got us. With our lunatic enthusiasm fomented by the exhilarating discovery that, Yes, By God, We Can Hear That Pesky Dither’nJitter(!), we decided that such audibility was in fact proof positive that Such Sound Sucks!

And this, when measured in the calm cool confines of the laboratory, turns out simply not to be true.

Based on the work I’ve done measuring the “audibility” of CODECs, where we use a 5-point rating scale (5 = no audible difference or “sounds exactly the same,” 4 = audible but not annoying difference or “sounds fine but I can tell them apart,” 3 = audible and slightly annoying difference or “doesn’t sound quite as good,” 2 = audible and distinctly annoying difference or “actually it sounds pretty bad” and 1 = audible and extremely annoying difference or “this sound sucks!”), we have found that a good CODEC can score above 4 (“audible but not annoying”) on average for expert listeners, in comparison to a 16-bit 44.1 kHz. signal. Such a finding is incompatible of course, with the assessment that “this sound sucks,” a score of 1. Meanwhile, as part of a controlling calculation in the test design, we compare CDs to themselves as part of such tests, and they score, as you would imagine, very close to 5, like 4.8. They don’t score 5, of course, because humans make mistakes. It’s very comforting.

Now then, I’ve listened to 24-bit audio and I’ve listened to 16-bit audio, and what I have to tell you is that the difference between 24-bit and 16-bit audio is a great deal smaller than the difference between 16-bit and the BEST CODEC I’ve measured as yet. Based on my lab experience, I can quite reasonably tell you that expert listeners, in blind test conditions, simply are not going to give 16-bit signals scores of less than 4 (“audible, but not annoying”), on average, on a reasonably scaled test. My expectation is that in such a test the average score for 16-bit in comparison with 24-bit audio would be around 4.3 to 4.6. I would expect a similar result comparing a 16-bit signal to an analog signal that hadn’t been recorded (a recorded analog signal will probably be audibly worse, due to the introduction of tape noise at about the level of the 14th bit – interesting, eh?).

This means, of course, that 16-bit audio sounds fine. It doesn’t introduce any significant problems into the analog signal. It may in fact be indistinguishable from 24-bit under these conditions. Even if not, it really does very well. You can relax.

Why all the fuss, then? Next month we’ll look at the forces that have driven this hi-rez process. Should be fun.

Thanks for listening.
NEXT> Why the fuss?    
1 2 3 4 5 6 >

COMMENTS

Russia     Aug 15, 2006 09:39 PM
Hi,
First i must say i'm no pro at recording, neither a pro musician. But as a physics studing i'm trying to apply some logic rather than to write 'by experience'.
I totally agree with you, but i think there are some issues that need more detail...

As you were talking about bit-depth in DSP processing, a thought came to my mind -- there are some effects(distortion, fuzz-es and such) that do introduce a huge gain boost like those 40dB needed to hear all the subtleties. It's not that i say that some sound source-sound detail will arise, but still the sound of the electric/thermal/acoustical noise now needs to be reproduced faithfully for the processed version not to be too-unnatural sounding. I think this is of utmost importance to some super-vintage-analog sounding-gear. So really, for such applications at least 20 bit source is a must. But it's all mostly about some "super-rectifer"-heavy metal gutars or some grunge etc and some other guitar gear.

And the second, as far as i know most modern DAW use intermediate 64-bit floating-point for mixing, or at least mix at 24 bit. I know that article dates 2001, but still. So source files don't really have to be more than 16 bit for this purpose, especially because they are usually multiplied by a less than one coefficient at a much higher resolution. So all the math issue just disappears. This can also be said about DSP processing -- no one said it has to be done at the same bit-depth as the source.
Victor Starodub 
canada     May 08, 2012 03:42 PM
I am a pro musician, and perform onstage as a live classical soloist.

I have found all these technical discussions quite interesting. I am not a technician, but I do own recording equipment and a high fidelity stereo system.

I recently made a recording using, at varying performances, 16 44, 24 48, and 24 96 settings. There did seem to be some small differences in the recordings, but the 16 44 was fine.

However, I find the hierarchy of need for a successful recording is in the following order::

1. The quality of the live performance
2.. The quality of the acoustics in the hall.
3. The position of the microphones.
4. My mood as I listen to the recording.
5. The skill of the sound people.
6. The quality of the equipment.

Some of my favorite recordings are of terrible technical quality and in mono, but of great artists.
ken 

Post a Comment



rss2

rss atom