Microsoft steadfastly refuses to reveal the implementation details behind its Windows Media Audio compression algorithm. Aficionados generally agree that this codec has good performance coupled, uncharacteristically, with fast encoding speeds.

Several independent listening tests—as well as my own so-far-unscientific experiences listening to WMA-encoded music and test tones—suggest that at bit rates one-third to one-half lower than the commonly employed 128- to 192-kbit/sec MP3 streams, its quality is at least as good as the industry-standard AAC algorithm, perhaps better. In the absence of official word from Redmond, audio enthusiasts such as myself are doing our best to figure out how WMA works its magic.

In a 1999 EDN article ("Digital audio breaks the sound barrier,"), I suggested that Microsoft might be employing a variant of the vector-quantization technique used by a lossy codec called TwinVQ. The TwinVQ encoder and decoder both contain a prefabricated set of data coefficients called a codebook, representing what the algorithm's developers believe to be the most common sets of per-frame frequency combinations found in audio. The TwinVQ encoder, after performing time-to-frequency transformation and compressing each frame of audio data, finds the closest match in its codebook and instead of sending the actual coefficient data, instead sends the codebook index. The decoder uses this index to pull the matching data approximation from its matching codebook, which it then outputs.

Since each codebook index is much smaller than the actual data, vector quantization can produce impressive file-size reductions. But there's a tradeoff. If the codebook is poorly constructed, or if the encoder doesn't do a good job of matching the actual data to a codebook entry, the compressed audio can sound terrible. Therefore, several variations on the basic vector-quantization technique are possible. Instead of using a preassembled codebook, the encoder might create it on the fly based on the specific characteristics of the audio it's compressing.

The advantage of this on-the-fly approach is that you're more likely to get good matches between data samples and codebook entries. However, now you have to transmit the codebook along with the compressed audio, because the decoder won't already have it. Therefore, the larger (and theoretically better quality) the codebook, the larger the compressed bit stream and the less efficient the compression results. Also, on-the-fly codebook creation is computing-intensive, leading to slow encoding speeds. As an interim step, therefore, the vector-quantization algorithm may rely mostly on a pre-fabricated codebook, supplementing it with a smaller codebook appendix created on the fly.

In response to that EDN article, Sean Alexander, Product Manager for Microsoft's digital media division, responded that "we don't do vector quantization." However, some reported characteristics of Windows Media Audio lead me to suspect that although Microsoft might not be employing a strictly defined vector-quantization approach, their algorithms might be analogous to, or derived from, techniques similar to TwinVQ. I've heard from several sources, for example, that short snippets of WMA-encoded audio sound better at a given bit rate than entire encoded songs. This feedback wouldn't make sense if, like MP3 and many other perceptual coders, WMA simply transformed and compressed several-millisecond long multisample frames in a one-at-a-time fashion.

Quality degradation with increasing audio duration could, however, occur if the encoder was creating a codebook on the fly. Longer audio sequences tend to contain more randomness than shorter clips—randomness that's less accurately approximated by a fixed-size codebook. A back-and-forth discussion on Internet newsgroup rec.audio.pro between Arny Krueger, maintainer of the PC ABX audio testing Web site, and Amir Majidimehr, Microsoft's general manager for digital media, also hints at this phenomenon (to find the thread, search on topic "Sound & Vision's 'Download Showdown' MP3 versus AAC & Windows Media"). Amir was unable to replicate Krueger's results, in which WMA-encoded versions of test clips exhibited pre-echo, quantization noise, missing information, and other artifacts. What Krueger and Majidimehr ended up figuring out was that whereas Majidimehr encoded each test clip separately, Krueger combined all of his test clips into one big file before running them through the WMA encoder.

Majidimehr wrote in one of the newsgroup postings that "Indeed, a combined input file of all the samples does generate different results than encoding each clip, which is what I did. If you want to gang encode a lot of samples, it is necessary to leave sufficient space between them [Editor's note: Later in the posting Majidimehr recommended 5 seconds] so that you truly get independent samples. Otherwise, your results only represent the composite, as the previous clip may change the outcome on some codecs such as ours." I hesitate, though, to absolutely conclude that WMA is using vector-quantization techniques, for two reasons. First is Alexander's "no vector quantization" comment, although something peculiar is definitely going on.

WMA also exhibits a curious discrepancy from TwinVQ and other similar vector-quantization approaches to compressing multimedia such as still images and video). Vector quantization has a reputation for being extremely slow, particularly when the algorithm incorporates on-the-fly codebook creation. WMA encoding, however, is reputed to be very fast, compared with other lossy compression routines at comparable bit rates. Perhaps Microsoft's engineers have just written tight code that takes advantage of processor acceleration hardware such as MMX instructions. I'll dig into the algorithm myself and report back any interesting results I see.

In September, the folks at Verance invited me to Sony Music Studios (Santa Monica, CA) for a listening test of their watermarking technology. Audio watermarking hides media identification bits within the audio bit stream, in a supposedly inaudible fashion.

A number of different watermarking alternatives exist, including spread-spectrum approaches that derive from the same mathematical theories behind CDMA digital cellular systems, and perceptual coding, which Verance employs. As the name implies, perceptual watermarking uses similar techniques to those in MP3, AAC and other perceptual audio-compression algorithms, with similar artifacts.

The half-hour audition Verance gave me was an abbreviated version of the testing they'd been conducting for a number of months with a large number of listeners as they fine-tuned their algorithms. I was able to choose from among several different music clips. Ten times I first listened, for as long as I wanted, to "clear" original audio (sample A), then watermarked audio (sample B), then to an unidentified sample X. I was then asked to determine whether sample X matched sample A or B. This testing method is known as ABX testing.

Qualifiers first. I knew that perceptual watermarking probably produced perceptual compression-like phenomena, for which my ears were already pre-tuned from my codec work of the past few months. I also intentionally picked an audio selection that would most accentuate these phenomena—a jazz passage with lots of high-frequency energy and sharp percussive transients.

Now the results. In all 10 cases, I was able to clearly identify differences between Samples A and B, such as less distinct, more muddled transients in the latter, as well as reduced high-frequency information and a narrower stereo image. And, more importantly, in 8 of 10 attempts I correctly matched Sample X to either Sample A or Sample B. This success rate translates to only a 5 percent probability that I was guessing. And I'd go so far as to say that, practically, the probability was even lower, because one of the two matches I missed was the 10th, when my accuracy was probably diminished due to listening fatigue.

Here's how I predict Verance will dispute my results. First, the company will claim that if I ran the experiment again, my success percentage might be lower. I have no way to dispute this possibility. But my success percentage might just as well improve.

Secondly, the company will claim that I was predisposed to hearing perceptual artifacts because of my previous experience auditioning lossy compression algorithms. This claim I can more readily dispute. In fact, much of the criticism directed to date at Verance centers on the fact that the company pulled people off the street and ran the tests without pre-training them. Audio experts pretty much agree that the types of defects created by lossy compression of both audio and video become more noticeable as consumers spend more time with the material. For example, I certainly notice MPEG video-compression artifacts more now than I did when I watched my first DVD.

Third, Verance will claim that the audio material I chose was worst-case and not indicative of results across the broad spectrum of music genres. Again, without a more extensive listening test I can neither refute nor support this viewpoint. I will say that the watermarking effects were far subtler than those created by, say, 128-kbit/sec MP3 compression. What's more, they would be masked fairly effectively by low-cost portable audio players, by the reduced dynamic range and frequency range of an automobile listening environment, and even by home stereos of less-than-audiophile quality. To Verance's credit, I'll also point out that in no case did Sample B sound "bad," and that in the absence of the clear Sample A as a reference, I'd probably be perfectly happy listening to Sample B.

But that's not the point. The DVD Audio discs with which Verance watermarking will debut target discriminating audiophiles. Even putting debatable reality aside, why would these folks pay thousands of dollars for equipment, and dozens of dollars per disc, for sound that they perceive to be altered by a watermark? I don't dispute the music industry's valid right to protect content they own from copyright infringement. But watermarking appears to be a technology that only benefits the sellers, not the buyers. And in fact, for the latter, it's a detriment. As such, from where I'm sitting it seems destined to fail.

Unlike other codecs, the ATRAC encode/decode algorithm used by MiniDisc equipment runs in dedicated hardware, not in software residing on a PC. Most MiniDisc units, such as my Sharp MT-MD15 portable player/recorder, provide digital inputs but not digital outputs. Without moving the data to my PC, I can't analyze it. But it would be unfair for me to subject the already-altered ATRAC audio to further degradation (digital-to-analog conversion coming out of the player's headphone or line output and analog-to-digital conversion coming into the PC's line input). So, for several weeks I regularly pestered the MiniDisc manufacturers (Kenwood, Sharp, Sony) to let me borrow a higher-end MiniDisc player/recorder with full digital I/O.

My efforts failed. As soon as I mentioned that I wanted to perform benchmarking against other codecs, responses to my emails and phone calls mysteriously dried up. Fine. Two can play that game. I went to eBay and bought a Kenwood MD-203 player/recorder. So there.

I'll be running the Kenwood's digital outputs directly into the digital inputs of my PC's sound card. I should also point out that I'll be testing version 4.5 of the original 292-kbit/sec ATRAC algorithm, not the newer 132-kbit/sec and 66-kbit/sec variants, variously known as ATRAC3 and MDLP (for MiniDisc Long Play). ATRAC3, developed in response to MP3 and other high-quality, low-bit-rate codecs, finds use in newer Sony MiniDisc players as well as the Memory Stick Walkman and Vaio Music Clip.