How it works

The input is the Weimar Jazz Database (WJazzD), release 2.1: 456 manually prepared monophonic solo transcriptions, with 200,809 note events in this import. We use each note’s rounded semitone pitch, onset and duration in seconds, and the database beat grid. Metadata supplies performer, instrument, composition, and recording date. The index also holds 1,102 piano-trio solos from the Jazz Trio Database (JTD), whose chordal piano is reduced to its top note per onset and whose beats come from its tracked beat grid; JTD solos play from the performance’s own YouTube video. Other prepared collections are not included in the published index.¹²

Onsets and note endpoints are interpolated onto the beat grid. For a solo with pitches p_i, the interval sequence is:

Interval sequences preserve the shape of an exact phrase under transposition. They do not make the search invariant to octave displacement or changed interval sizes.

1. Seed. Index every five consecutive pitch intervals, spanning six notes. Occurrences of the same seed in different solos form candidate anchors. Seeds with more than 80 occurrences are omitted to limit very common patterns.
2. Extend. From each anchor, search rightward. A step may pair the next notes or skip one note on either side. The seed sets a fixed transposition; subsequent paired notes pay a pitch and rhythm cost, while each skip costs 1.
3. Select. Keep the longest reachable endpoint meeting the constraints below, with at least eight paired notes. Then remove contained windows. Within one ordered solo pair, windows overlapping by at least 50% of the shorter range on both solos are consolidated, keeping the higher-evidence result.

This is an exhaustive search over eligible interval seeds with a bounded rightward alignment. It is a heuristic devised for this archive, not an implementation or validation of the pattern-search systems described in related Jazzomat work.²

For each paired note, let P be transposition-adjusted pitch difference capped at 1 after division by two semitones. Let O and D be differences in inter-onset spacing and duration, both in beats: each has a one-eighth-beat tolerance, then rises to 1 over a further half beat. The paired-note cost is:

The total cost C includes these paired-note costs and one unit per skipped note. Search stops a path when C exceeds 6, or cost accumulated since the last zero-cost pair exceeds 2. A qualifying passage has:

Here N is the number of paired notes and G the number of skips. The default “Best matches” order uses evidence N − C; it is a ranking heuristic, not a probability of quotation. All search and consolidation run before publication; the site reads the resulting static catalog.

A match establishes similarity between two symbolic passages in this corpus. It does not by itself establish borrowing, influence, or priority. Shared repertoire, idiom, transcription choices, and chance are alternative explanations. Recording dates come from corpus metadata. Linked recordings, when available, use separate video alignment data and may have imprecise timestamps.³

In the example, the shared core is a mostly chromatic descent. The two recording dates place Parker’s solo earlier, but the algorithm cannot determine whether Coltrane drew on that performance or on common jazz vocabulary.

The exact five-interval seed excludes short or heavily transformed motifs; the frequency cap omits some common material. Rightward extension can begin after a musically meaningful start. Pitch and rhythm weights, the 0.80 floor, and the 50% consolidation threshold were chosen heuristically and have not been calibrated against expert relevance judgements. The result set is therefore a discovery aid, not a recall-complete or statistically validated detector.