i Tunes store, the profit margins on delivering rich media content are enough to induce sweaty palms in any chief executive.

AT&T has been rolling out their cable TV competitor, U-Verse, which delivers media content over the same high-capacity lines that also carry Internet data. If they smell money to be made offering exclusive media, attracting partnerships with record labels and studios could be reason to embrace the kind of content filters these companies are so desperately seeking. But what about safe harbor and liability? Pundits are still scratching their heads.

Digital Details

Talking about content filtering is one thing—actually doing it is quite another. An effective filter needs to identify audio or video data that could be contained in a wide variety of formats. Data can be compressed, transformed, and distorted in a wide variety of ways. It is not difficult for a human to recognize that a camcorder recording of a theatrical screening is the same content as the DVD rip. It is much more difficult for a computer to do so.

Furthermore, an effective filter needs to make its distinctions quickly, using the smallest samples of data feasible, with the least computation demands possible, or else it will exact a costly performance penalty on the whole network.

Back when the original Napster was spooking the music industry, record labels looked for infringing songs with a very

primitive approach: by filename. Wily file sharers simply changed the names of files to misidentify them. Today’s filters are a lot more sophisticated. But even after years and millions of dollars of development, they remain imperfect.

Acoustic fingerprinting like that developed by GraceNote is designed to calculate a unique “score” for a particular audio track based on its perceptual qualities. Compression techniques like MP3 and AAC use psychoacoustic models to determine which audio bits can be “thrown away” while still maintaining a similar perception experience for the listener. Likewise, acoustic fingerprints have to go beyond the exact set of bits that make up a particular file, because every compressor will produce a slightly different file.

The technology behind acoustic fingerprinting is now quite effective. But to identify any individual track, you need a fingerprint to compare it to, and that means a database—a huge database. For the database to be useful, it needs to contain millions of acoustic fingerprints calculated from authorized recordings.

Such a database can never be truly comprehensive—a live recording of a U2 concert may not produce the same acoustic fingerprint as the tracks from their CD, for example. Plus, GraceNote is not alone in licensing fingerprinting services, meaning that multiple vendors with different algorithms compete in the market. It is not realistic to expect every record label to pro-