@audio/mir-fingerprint

Landmark audio fingerprinting (Wang 2003, Shazam class) — spectrogram peak constellation → anchor→target hashes (f1, f2, Δt) → matching by offset-histogram vote. Robust to noise and level; exact-match identification, not similarity.

npm install @audio/mir-fingerprint

import fingerprint, { match } from '@audio/mir-fingerprint'

let fp = fingerprint(data, { fs: 44100 })
// Array<{ hash: number, t: number }> — t in frames (frame = hop/fs seconds)

let r = match(fpA, fpB)
// { score: number (votes for the best time offset), offset: number (frames, b within a) }

fingerprint() finds prominent spectral peaks per frame (magnitude > 4× the frame mean and a local max), keeps the strongest peaksPerFrame, then pairs each peak as an anchor with up to fanout forward peaks within window frames — each pair becomes one hash landmark.

match() builds a hash→time index from fpA, then for every fpB landmark with a matching hash accumulates a histogram of time offsets; the offset with the most votes is the alignment, and its vote count is the match confidence. A near-duplicate clip scores high at one dominant offset; unrelated audio spreads votes thinly across many offsets.

Options: - fs — sample rate (default 44100, Hz) · frameSize — analysis window, samples (default 1024) · hop — frame hop, samples (default 512) · peaksPerFrame — kept peaks per frame (default 5) · fanout — target peaks paired per anchor (default 5) · window — max anchor→target Δt, frames (default 32)

Also exported as an audio.js stat manifest (./audio).

Part of @audio/mir.