Shazam: a song recognised from a snippet
At ISMIR 2003 in Baltimore (26-30 October 2003) Avery Wang of Shazam Entertainment described an algorithm that recognises a recording from a few seconds of sound picked up by a phone's microphone. At the time of the paper the service held over 1.7 million tracks and was live in the UK and Germany; in the UK it had launched in August 2002, reached by dialling 2580.
Why it matters
Music heard through ambient noise became a query to a database: hold a phone up to a speaker and the answer arrives as a text message. The algorithm keeps not the whole sound but the positions of a spectrogram's peaks and pairs of those peaks, so a search among a million tracks takes milliseconds and survives voices, street noise and codec compression.
According to the paper, the algorithm works below -6 dB signal-to-noise ratio for white noise and down to -9 dB when someone speaks over the music; from a heavily corrupted 15-second sample a match is found even when only 1-2% of the hashes survive. On about 20,000 tracks on a single PC a search takes 5 to 500 milliseconds. Each hash is a 32-bit number built from the two frequencies of a pair of peaks and the time between them. The service took up to 15 seconds of audio and replied by SMS with title and artist; the company was founded in 2000. Telecompaper reported the launch: on 11 June 2002 that Vodafone, Orange, O2 and T-Mobile were preparing the service for launch in July or August, and on 16 September that Shazam had launched in August, that a call to 2580 cost up to 50 pence a song and that the database held over 1.5 million songs. What this record does not assert: every figure in the paper comes from the company's description of its own system and no independent test was found, hence medium confidence. The 2002 reports do not say whether the service went live on all four networks at once. No one names the day of the launch: 19 August is the date of Apple's 2022 update, not of the launch.