If you are an audiophile who uses a PC as a source in your audio system, you’re probably aware of the fact that Windows Vista introduced a brand-new audio engine to replace the much hated KMixer of Windows XP. In my opinion, there are a few reasons why audiophiles should be happy with this change:

• The new audio stack automatically upconverts all streams to a 32-bit floating-point sample depth (the same that is used in professional studios) and mixes them with the same precision. Because of the amount of headroom that comes with using 32-bit floats, there is no more clipping when playing two samples at the same time. There is also no loss of resolution when you lower the volume of a stream (see below).
• The Vista/Win7 audio engine automatically feeds your sound card with the highest-quality output stream that it can handle, which is usually 24 bits per sample. Perhaps you’re wondering why you should care, given that most music uses only 16 bits per sample. Suppose you’re playing a 16-bit song with a digital volume control set to 10%. This corresponds to dividing each sample by 10. Now let’s assume the song contains the following two adjacent samples: 41 and 48. In an ideal world, after the volume control we would get 4.1 and 4.8. However, if the output stream has a 16-bit depth just like the input stream, then both output samples will have to be truncated to 4. There is now no difference between the two samples, which means we have lost some resolution. But if we can have an output stream with 24 bits per sample, for each 16-bit level we get 2⁸ = 256 additional (“fractional”) levels, so we can still preserve the difference between the two attenuated samples. In fact, we can have ≈4.1016 and ≈4.8008, which is within 0.04% of the “ideal” samples of 4.1 and 4.8.
• Don’t you hate it when you change the volume in your movie player or instant messaging software and instead of changing its own volume, it changes your system volume? Or have you ever used an application with its own poorly implemented volume control (iTunes, I’m pointing at you!)? Well, these abominations should now be behind us. In Vista and Win7, each application gets its own audio stream (or streams) and a separate high-quality volume control, so there should no longer be any reason for application vendors to mess with the system volume or roll their own and botch the job.

So Windows Vista and Windows 7 upconvert all your samples to 32-bit floats and mix them with 32-bit precision into an output stream that, by default, has the highest bit depth that your hardware can handle. The output bit depth is customizable; you can change it in the properties of your audio device. If you change it e.g. to 16 bits, the audio engine will still use 32-bit floats for internal processing — it will just downconvert the resulting stream to 16 bits before sending it to your device.

Now, what about the sample rate? You can set the output sample rate in the audio device properties window, but is there also some internal sample rate that the Windows audio engine uses regardless of your setting? For example, does it upsample your 44.1 kHz songs to 96 or 128 kHz? Unlike the upconverting from 16-bit integers to 32-bit floats (which should be completely lossless), this could potentially introduce some distortion as going from 44.1 kHz to 96 or 128 kHz requires at least some interpolation.

I couldn’t find the answer to this question anywhere, so I wrote to Larry Osterman, who developed the Vista and Win7 audio stacks at Microsoft. His answer was that the sample rate that the engine uses is the one that the user specifies in the Properties window. The default sample rate is chosen by the audio driver (44.1 kHz on most devices). So if your music has a sample rate of 44.1 kHz, you can choose that setting and no sample rate conversion will take place. (Of course, any 48 kHz and higher samples will then be downsampled to 44.1 kHz.)

There is some interesting technical information on the Windows Vista audio stack in this Channel9 video.

Policy no. 8 in the Terms of Service of the respected audiophile community Hydrogenaudio states:

8. All members that put forth a statement concerning subjective sound quality, must — to the best of their ability — provide objective support for their claims. Acceptable means of support are double blind listening tests (ABX or ABC/HR) demonstrating that the member can discern a difference perceptually, together with a test sample to allow others to reproduce their findings.

What a breath of fresh air. Other audio forums are full of snake-oil-peddling and Kool-Aid-drinking evangelists who go on and on about how replacing $200 speaker wires with $400 speaker wires “really opened up the soundstage and made the upper-midrange come alive”. The people at Hydrogenaudio know that such claims demand proper scientific evidence. How nice to see that they dismiss subjective nonsense and rely instead on the ultimate authority of ABX tests, which really tell us what makes a difference and what doesn’t.

Except that ABX tests don’t measure what really matters to us. ABX tests tell us whether we can hear a difference between A and B. What we really want to know, however, is whether A is as good as B.

1.

“Wait a second!”, I hear you exclaim. “Surely if I cannot tell A from B, then for all intents and purposes, A is as good as B and vice versa. If you can’t see the difference, why pay more?”

Actually, there could be tons of reasons. To take a somewhat contrived example, suppose I magically replaced the body of your car with one that were less resistant to corrosion, leaving all the other features of your vehicle intact. Looking at the car and driving it, you would not notice any difference. Even if I gave you a chance to choose between your original car and the doctored one, they would seem identical to you and you could choose either of them. However, if you were to choose the one I tampered with, five years later your vehicle’s body would be covered in spots of rust.

The obvious lesson here is that “not seeing a difference” does not guarantee that A is as good as B. Choosing one thing over another can have consequences that are hard to detect in a test because they are delayed, subtle, or so odd-ball that no one even thinks to record them during the test.

But how is this relevant to listening tests? Assuming that music affects us through our hearing, how could we be affected by differences that we cannot hear?

In his fascinating book Burning House: Unlocking the Mysteries of the Brain, Jay Ingram describes the case of a 49-year-old woman suffering from a condition called hemispatial neglect (the case was researched by neuropsychologists John Marshall and Peter Halligan). Patients with hemispatial neglect are unable to perceive one (usually the left) side of the objects they see. When asked to copy drawings, they draw only one side; when reading out words, they read them only in half (e.g. they read simile as mile).

In Marshall and Halligan’s experiment, the woman was given two simple drawings showing two houses. In one of the drawings, the left side of the house was covered in flames and smoke; the houses looked the same otherwise. Since the flames were located on the left side, the patient was unable to see them and claimed to see no difference between the drawings. When Marshall and Halligan asked her which of the houses she would rather live in, she replied — rather unsurprisingly — that it was a silly question, given that the houses were identical.

However, when the experimenters persuaded her to make a choice anyway, she picked the flameless house 14 out of 17 times, all the time insisting that both houses look the same.

Marshall and Halligan’s experiment shows (as do other well-known psychological experiments, including those pertaining to subliminal messages) that it is possible for information to be in a part of the brain where it is inaccessible to conscious processes. This information can influence one’s state of mind and even take part in decision-making processes without one realizing it.

If people can be affected by information that they don’t even know is there, then who says they cannot be affected by inaudible differences between an MP3 and a CD? Failing an ABX test tells you that you are unable to consciously tell the difference between two music samples. It does not mean that the information isn’t in your brain somewhere — it just means that your conscious processes cannot access it.

So the fact that you cannot tell the difference between an MP3 and a CD in an ABX test does not mean that an MP3 is as good as a CD. Who knows? Maybe listening to MP3s causes more fatigue in the long run. Maybe it makes you get bored with your music more quickly. Or maybe the opposite is true and MP3s are actually better. We can formulate and test all sorts of plausible hypotheses — the point is, an ABX test which shows no audible difference is not the end of the discussion.

2.

I have shown that the lack of audible differences between A and B in an ABX test does not imply that A is as good as B. Before you read this post as an apology for lossless audio formats, here is a statement that will surely upset hard-core audiophiles:

The fact that you can tell the difference between an MP3 and a CD in an ABX test does not mean that the MP3 is worse than a CD.

First of all, the differences between MP3s encoded at mainstream bitrates (128 kbps and 192 kbps) and original recordings are really subtle and can be detected only under special conditions (quiet environment, good equipment, full listener concentration, direct comparisons of short samples). Because the differences are so tiny, we cannot automatically assume that it is the uncompressed version that sounds better. Subtle compression artifacts such as slightly reduced sharpness of attacks on short, loud sounds may in fact be preferred by some listeners in a direct comparison.

Secondly, even if we found that the uncompressed version is preferred by listeners, that wouldn’t necessarily mean that it is better. People prefer sitting in front of the TV to exercising, but the latter might make them feel much better overall. If it were discovered, for example, that compressed music is less tiring to listen to (this is of course pure speculation), then that fact might outweigh any preference for uncompressed sound in blind tests.

Summary

The relevance of ABX tests to the lives of music lovers is questionable. Neither does the absence of audible differences imply equal quality, nor does the presence of audible differences imply that the compressed version is inferior. Rather than being the argument to end all debate, the results of ABX tests are just one data point and the relative strengths of various audio formats may well be put in a new light by further research.