How Does Shazam Actually Work?

The sound picked up by a phone's microphone is stored as a waveform, but it's hard to compare directly once volume and playback environment change. Running FFT over short segments breaks the sound into frequency components, which are stacked to build a spectrogram.

• Only the strongest peaks in the spectrogram are kept, compressing it into a sparse set of points.
• Most background noise is discarded, leaving only the dominant frequency landmarks.
• This process runs on top of standard audio input, such as 44,100 Hz sampling.

Each peak becomes an anchor, which is paired with peaks inside a target zone to its right, creating a hash from the two frequencies and their time difference. Even a 3-minute song can produce thousands of these fingerprint hashes, and the result ends up being closer to a specific recording than to the song as a whole.

Instead of scanning songs one by one, the search uses a hash-first inverted index. Hashes obtained from the clip are looked up directly, and matches are further checked against the time offsets between matching hashes to reduce false positives.

• Server-based Shazam uses a large-scale fingerprint database,
• while Apple's on-device recognition and Google Pixel's Now Playing use a smaller local DB with ML optimization.
• The key isn't storing a lot of information, but precisely discarding the information that can be discarded.

The foundation is Avery Wang's 2003 paper An Industrial-Strength Audio Search Algorithm.