how the shazam algorithm works & how i cloned it

i built a song-recognition algorithm from scratch this summer, and the part that stuck with me wasn’t the fourier math — it was the indexing.

the naive version is unusably slow

the first version matched query fingerprints against the database point by point. correct, but hopeless at any real scale.

the fix was a combinatorial hashing scheme: instead of hashing single spectral peaks, hash pairs of peaks along with the time delta between them. suddenly each fingerprint is far more distinctive, and lookups become a hash-table hit instead of a scan — about a 10,000× speedup in query time.

recognition held up in noise

using time-offset histogram matching, the recognizer hit ~90% accuracy at signal-to-noise ratios above +7 dB. the histogram trick is elegant: correct matches all agree on the same time offset, so they pile up into one tall bin while noise scatters flat.

there’s a longer write-up coming, but the short version: the interesting engineering was never the transform. it was making the search cheap.