The encyclopedia · Software & IT · Technical decision · 2000–2003
Shazam matched songs by hashing their fingerprint, not by listening.
Instead of trying to understand a song, Shazam turned audio into robust landmark hashes so a short, noisy phone clip could identify the track out of millions.
Shazam Entertainment
the move
The idea behind Shazam, started in 2000, was to let a phone identify the music playing around it. The algorithm had to recognize a short clip that had been broadcast, mixed with heavy ambient noise, squeezed through a tiny phone microphone and voice codec, and dropped in transit, all against a library of millions of tracks.
Avery Wang's solution was to stop trying to understand the audio and instead fingerprint it. The system scans the spectrogram for local energy peaks and pairs nearby peaks into landmark hashes. Matching those hashes against a database of over a million tracks finds the song, and because the hashes of a matching track must be linearly related in time, false positives stay rare.
The ISMIR 2003 paper reported recognition robust to foreground speech, distortion and compression, with search times of a few milliseconds per query for radio monitoring. It converted music identification from a transcription problem into a matching problem, which is exactly what made it fast.
why it works
- Energy peaks survive noise, so a noisy recording maps to the same hashes as the clean track.
- Combinatorial hashing lets a tiny clip be matched across a database of over a million tracks.
- Recognizing features beats recognizing content, so no transcription or song metadata is needed.
- Requiring matched hashes to align in time kills most false positives.
what transfers
When the full signal is too noisy to recognize, extract the few stable features that survive the noise and match on those instead of trying to parse the whole thing.
what came after
Shazam became the default way people name music, and the landmark-hash approach now underpins content ID, broadcast monitoring and audio copyright systems. The algorithm from a 2000 startup still powers an app that handles billions of recognitions.
references
- An Industrial-Strength Audio Search Algorithm (Avery Li-Chun Wang)
- System and methods for recognizing sound and music signals in high noise and distortion (U.S. Patent 6,990,453)
spotted an error? The archive wants to know.