Skip to main content

What CORPUS Knows About Each Track

Public beta

CORPUS is being built in the open. Some of what you read here is live, some is still design intent — expect it to evolve.

A search engine is only as expressive as the dimensions it indexes. The dimensions below are what every track in the corpus carries, and what every query — precision, similarity, narrative — resolves against.

Acoustic and structural​

The properties extracted directly from the audio signal:

  • Tempo (BPM), mapped to the conventional tempo vocabulary (Larghissimo through Prestissimo) so queries can read naturally — "around Andante", "faster than 140 BPM".
  • Key, as tonic plus mode (major, minor, modal). Enharmonic equivalents are expanded automatically in search: a query for C♯ minor also matches D♭ minor.
  • Energy, normalised so that "low-energy ambient" and "high-energy punk" are operational categories, not figures of speech.
  • Instrumentation, as a multi-tag set — strings, synths, drums, piano, brass, and so on, at sufficient resolution to support exclusion ("no piano") and inclusion ("prepared piano specifically").
  • Vocal treatment — instrumental, with vocals, and where vocals are present, the vocal character (male, female, mixed). Personality-rights consent at upload is what makes the with vocals path usable at all; see Ownership and Consent.

Semantic and contextual​

The properties that describe the work rather than the signal:

  • Genre and sub-genre, drawn from a catalogue vocabulary that grows with the corpus and is reconciled at ingest.
  • Mood, as a multi-tag set — dark, uplifting, contemplative, anxious, and so on. Mood is treated as orthogonal to genre, because two tracks that share a genre can do very different emotional work.
  • Keywords, freetext tags from contributor annotations and from the pipeline itself.
  • Provenance, recording the source repertoire a track came in from. Useful for filtering, useful for audit.

Narrative and dramaturgical​

The properties used by Story2Music to reason about how a track functions in a scene:

  • Narrative function — what role the track plays (build-up, release, transition, undercurrent).
  • Emotional arc — how the affect changes across the work, not just what it averages to.
  • Dramaturgical fit — the dimensions on which a piece can transform a scene (tension, intimacy, propulsion, distance).

These are not surface tags. They are the layer that lets the system propose three different musical interpretations of the same scene rather than three variations of one.

Originality and novelty​

Two signals that feed both scoring and search:

  • Originality, measuring how a track positions itself relative to the existing library — the same relational analysis that drives the diversity dimension of scoring.
  • Novelty, capturing structural or stylistic distinctiveness independently of catalogue density.

What is still being built out​

A few dimensions are in active development rather than fully shipping:

  • Production quality assessment. Early results are promising at identifying recordings of high production quality; the full bonus does not yet ship.
  • Cultural-context tagging. Beyond genre, the corpus is beginning to capture regional and tradition-specific markers that matter for therapeutic and culturally-sensitive applications.

The annotation layer evolves with the corpus and the applications it serves. New dimensions are added as new use cases require them; existing dimensions are recalibrated when the corpus grows enough that the old calibration no longer fits.