Audio Fingerprinting " A New Technology To Identify Music "

This report describes the core of the audio fingerprinting technology that was developed at Philips Research. An audio fingerprint is a unique code that uniquely identifies a segment of music like a human fingerprint identifies a human being. Audio fingerprinting can be used for several applications: • Broadcast monitoring to automatically generate playlists of radio stations • Connected audio: users can for instance identify a song by pushing a button on their radio or by dialing a service number and letting their mobile phone listen to the song. • Filtering technology for P2P networks: songs on the network can be identified irrespective of their metadata. Therefore copyrighted songs can for instance be blocked from the network. • Automatic Music Library Organization: digital music archives with inconsistent, missing or erroneous meta-data can be automatically labeled with the correct meta-data An audio fingerprint system basically consists of two parts: a fingerprint extraction algorithm and an algorithm to efficiently search such a fingerprint in a fingerprint database containing the fingerprints of many recordings. Fingerprint extraction is based on extracting bit sequences of 32 bits, so called sub-fingerprints, for every 11.8 ms of audio. These sub-fingerprints are obtained from energy differences along both the time and the frequency axis. A sequence of 256 sub-fingerprints (a fingerprint block), which corresponds to 3 seconds of audio, contains enough information to enable reliable identification. The objective of the fingerprint search algorithm is to find as efficiently as possible a matching fingerprint block in the fingerprint database. This is not a trivial task, because a fingerprint block will in general not have a bit exact copy in the database. This is due to the fact that a fingerprint block that has to be identified is generally taken from an audio excerpt that has been processed (intentionally or unintentionally). Therefore a novel algorithm is devised, which uses the fact that sub-fingerprints sometimes do not contain any bit errors. In case they do, a more sophisticated algorithm can be used, which uses reliability information of the extracted fingerprint bits.