- Raw PCM extraction: We parse the RIFF/WAVE header ourselves (no external libraries) to load 16 bit PCM samples.
- Spectrogram (STFT): Audio is framed into 2048-sample windows (≈46 ms) with 50% overlap, a Hann window is applied, and a radix-2 FFT is run to produce a high-resolution magnitude spectrogram (S[t,f]).
- Local maxima: For each time frame and frequency bin, we detect peaks that exceed a dynamic threshold (mean + 1.5 σ) and are the highest in a 7×7 neighborhood.
- Node list: Each peak (p_i = (t_i, f_i, A_i)) becomes a node in our fingerprint graph.
- Neighbor graph: For each peak (p_i), we connect it to up to K = 5 future peaks (p_j) within ±Δf (e.g. ±50 bins) and Δt (up to 1 s).
- Amplitude weighting: Each edge is weighted by (A_i \times A_j), so louder events carry more significance.
- Local descriptor: We record ((\Delta t, \Delta f, w)) for each of the K strongest edges, forming a 3×K vector around each anchor.
- Bins: (\Delta t) quantized into 16 time-bins, (\Delta f) into 32 freq-bins, and (w) into 8 amplitude-bins.
- FNV-1a hashing: The 3K bytes are concatenated and hashed into a 64-bit fingerprint code, compactly encoding subgraph structure.
- All
(code, songId, timestamp)tuples are bulk‐inserted into SQLite with an index oncode. - This on-disk database scales to millions of fingerprints and requires no external key-value service.
- Fingerprint snippet: The same STFT → peaks → graph → hashes pipeline runs on the query.
- Lookup: Each code is looked up in SQLite, returning
(songId, dbTimestamp)entries. - Voting: We tally votes for each
(songId, Δt = queryTime – dbTime)pair. - Best match: The song and time-offset with the highest vote count wins.
- Richer descriptors: Captures a small subgraph of K edges per anchor vs. a single peak pair.
- Amplitude-aware: Weighs edges by peak magnitudes for robustness to noise.
- Finer quantization: More granular time/frequency bins for higher discrimination.