Print

Print


Instead of a D.J., a Web Server Names That Tune
http://www.nytimes.com/2003/01/23/technology/circuits/23next.html?pagewanted
=all&position=bottom
WHAT'S NEXT

Instead of a D.J., a Web Server Names That Tune

   By ANNE EISENBERG

   A SONG and its name are easily parted, as anyone knows who has
   listened to a tune on the radio and then waited in vain for the
   announcer to identify it.

   Now some companies are using a technology that can name that tune as
   it plays, promptly displaying words like "Take Five by the Dave
   Brubeck Quartet" in text on an Internet radio or cellphone.

   The systems are sensitive enough to identify not only names and
   artists within a vast range of recorded music, but also different
   versions of a piece done by the performers, even when the differences
   are slight.

   At the recent Consumer Electronics Show in Las Vegas, for instance,
   Royal Philips Electronics demonstrated a prototype of an Internet
   radio that was not only capable of naming the band Pearl Jam as its
   music streamed past but also distinguishing a version of a tune that
   it played at a concert in Verona, Italy, from the same tune recorded
   in Milan.

   "To our ears the two versions sound the same," said Ton Kalker, a
   mathematician at Philips Research in Eindhoven, the Netherlands, who
   led the team that developed the rapid identification system. "But the
   technology is sensitive enough to make a distinction."

   The technology that spots the mystery tune in products from Philips
   and other companies is called audio fingerprinting. It is based on the
   premise that every performance of a song has unique audio
   characteristics - for example, a certain relationship of neighboring
   high and low notes over a minuscule slice of time. Represent those
   relationships in numbers, and you have a code that shows a particular
   version of a song, and no other.

   "Audio fingerprinting works by creating a mathematical description of
   some of the unique features of a song," said Dr. Richard Gooch, deputy
   director of technology at the International Federation of the
   Phonographic Industry, a trade organization based in London. The
   fingerprints are stored on a server. When the server is asked to
   identify a tune - for instance, a song playing on Internet radio -
   software matches a snippet of the tune, expressed in code, with the
   whole coded version of the song stored on the server.

   Some companies offer fingerprinting technology to identify not only
   streaming content but the contents of audio files and traditional
   radio broadcasts as well, Dr. Gooch said. His group and the Recording
   Industry Association of America have recently investigated many audio
   fingerprinting systems to see if they might benefit the recording
   industry.

   Dr. Gooch said that the technology stood up even to the most difficult
   conditions - poor loudspeakers, highly compressed streaming files or
   broadcasters who speed up songs slightly to make room for commercials.

   "We've found that even when broadcasters tweak a song or compress it,"
   he said, "so long as you can still hear it, the systems can extract a
   description of unique characteristics in the song," quickly matching
   the description with the database to identify the track.

   The steadily improving technology has been used commercially in
   business applications since at least the early 80's, he said, for
   example, in the music business to identify broadcast performances and
   then pay royalties to rights holders. Given the popularity of digital
   music, consumer applications of the technology are probably
   inevitable, Dr. Gooch said.

   "Audio fingerprinting is accurate, robust and runs in a sensible
   amount of time," he said. "It really works."

   Shazam Entertainment, based in London, already offers British
   consumers an audio fingerprinting service linked to cellphones. Users
   dial the service and then hold their phones up to the tune that is
   playing, say, on the car radio. Shazam has a database with
   fingerprints of 1.6 million tunes. It matches the incoming fingerprint
   with its database and within 30 seconds sends a text message back to
   the phone identifying the song.

   "We have most any popular song that has been recorded," said Philip
   Inghelbrecht, founder of the company and director of its content. "So
   long as the CD is commercially available, we will have it."

   In the future, manufacturers of electronic devices may offer audio
   fingerprinting to people who want to organize music collections stored
   on their hard drives, said Vance Ikezoye, chief executive of Audible
   Magic, a company in Los Gatos, Calif., that offers its own patented
   audio fingerprinting technology. "It's hard to manage music if you
   don't have the correct information for every song in your collection,"
   he said.

   Companies have used different techniques to create the unique code
   that constitutes a fingerprint. At Philips, for example, Dr. Kalker
   said, he and colleagues created the code by breaking each tune into
   10-millisecond snippets. Then they calculated the differences in the
   loudness of adjacent frequencies in the snippet and how those
   differences changed over time. They repeated the process every 10
   milliseconds to extract code over the entire length of the song.

   Then, when an Internet radio detected the song streaming through it,
   it sampled the song and sent the sample along for pattern-matching.
   "The whole process takes about three seconds," he said. "So even a
   dumb AM or FM receiver can tell you a lot about what you are listening
   to."

   Dr. Gooch suggested that fingerprinting systems would eventually be
   common on car radios. "When every car has its own digital audio
   player," he said, "people will want to know the name of the song they
   are listening to displayed on the dashboard."

   Mr. Ikezoye of Audible Magic said that people who used the company's
   system need not worry about violating the rights of music publishers.
   Although Audible Magic maintains a big database of popular copyrighted
   music, he said, the fingerprints are not the actual songs, but
   summaries of factual measurements describing the sound. "The original
   sound can't be reconstructed from the fingerprint," he said, "so
   storing and sharing fingerprints does not represent an infringement of
   the copyright."

   If the technology catches on in consumer applications, radio listeners
   may one day have an antidote to a common problem cited by Cary
   Sherman, president and general counsel of the Recording Industry
   Association of America: "Radios often don't bother to tell listeners
   what they have played," he said simply.