How we measure

Every number in the directory is either measured from audio or attributed to the analysis that produced it. This page explains where the data for 441 artist profiles and 66,923 tracks comes from, and what the provenance line on each track page means.

Catalogue

Artist catalogues, release groups, ISRCs and artist links are seeded from MusicBrainz and Discogs (both CC0 open data), then resolved against streaming catalogues for cover art, track ids and streaming links. Release-level genres and labels come from Discogs.

Key, BPM and audio features

Tempo, key, energy, danceability and the related features come from one of three analysis sources, tracked per track:

  • Model-based audio analysis: tempo, key and feature estimates from large-scale industry analysis models, matched per track. Most of the corpus.
  • AcousticBrainz: the open (CC0) Essentia-based analysis project. Used where a recording has a strong match.
  • Vibes measurement: our own signal processing over 30-second audio previews. This is what produces the waveform, the 96-band frequency spectrum and the dynamics (crest factor) figures, and it cross-checks reported tempo and key.

Each track page states which source its numbers come from, directly under the stats. Where sources disagree, we prefer the measured value and re-queue the track for analysis rather than average conflicting numbers.

Camelot and key notation

Musical keys are shown in standard notation and as Camelot wheel codes (8A, 5B, ...). The mapping between the two is deterministic; harmonic compatibility on track pages follows standard Camelot adjacency rules, with each move's energy effect labelled.

Waveforms, spectrum and dynamics

The waveform strip is a 200-point RMS envelope of the track's 30-second preview, measured by us. The frequency spectrum is a 96-band FFT average over the same preview. Dynamics is the crest factor in dB, a measure of how compressed the master is. These are measurements of the preview excerpt, not the full master, and are labelled as such on the page.

Generated text

Track and artist descriptions are computed from the measured data. A sentence only appears when the underlying number is statistically notable for its cohort (artist, genre or the whole directory), so two tracks in the same genre carry different text because their data differs, not because words were shuffled.

Corrections

Wrong key, wrong BPM, wrong artist resolve. If you spot one, tell us with the track link. Corrections go into the pipeline as overrides, so they survive every rebuild.