Case Study: A Harp Catalogue

musiscape was developed on a real corpus: the five-album, 57-track catalogue (2009–2023) of a solo harpist whose work spans melodic pedal-harp music, world-music colors, acoustic drone/meditation music, and reconstructions of ancient instruments (lyres, arched harps, monochords). A deliberately hard test: mostly rubato (beat trackers fail), mostly one instrument family (embeddings under-discriminate), with the interesting variance in timbre, texture and typology.

What the toolbox found — and verified

• Two sound-worlds. The corpus splits cleanly into melodic plucked albums (dense, dark, harmonic) and textural/drone albums (sparse, bright, inharmonic, wide dynamics) — visible in the fingerprints, the PCA landscape, and the affinity matrix alike.
• Internal consistency is measurable. The one solo-instrument album scored +0.51 self-similarity; an eclectic album scored +0.06. The same album's tracks also cluster tightly on the circle of fifths (R = 0.82 — neighbouring flat keys, harmonically a suite), while others wander (R ≈ 0.2).
• Categories found the outliers. k = 2 clustering didn't just reproduce album boundaries: it pulled the whistle, bodhrán and ensemble tracks out of their melodic albums into the "bright/textural" category — exactly the tracks with special instrumentation.
• The whole œuvre is quantifiably rubato: no instrumental track exceeded pulse clarity R = 0.22; the steadiest track (R = 0.28) was the one sung, strummed song. Drifting drones sat at R ≤ 0.005.
• Typology matched instrumentation: iterated tonic objects peaked on the monochord album, variable-mass objects on the album with buzzing lyre reconstructions, and impulse-vs-held shares separated plucked from bowed/drone facture.

Method notes

Extraction ran at 22.05 kHz mono over MP3 sources; findings were cross-checked against hand-rolled analyses (independent code paths) and against the CD booklets' instrumentation credits. Two real bugs were found by the corpus: the metrical-octave phase-cancellation trap in pulse clarity, and spurious onsets on featureless signals — both now fixed and regression-tested.