Methodology and Caveats

musiscape's core commitment: every number carries a musical name, and every method states where it stops working. This page collects the honesty notes in one place.

Interpretable features vs. embeddings

The descriptor set (note density, brightness, inharmonic texture, dynamics, pitch-class statistics, pulse clarity, Schaeffer proxies) is a deliberate trade against embedding models: weaker raw similarity, but every axis can be argued about. If a category or a landscape position surprises you, you can trace it to specific, auditable measurements.

Key estimation

Krumhansl–Schmuckler correlation on harmonic chroma. Reliable for tonal material; indicative only for drones (a single sustained pitch class correlates with several key profiles at once).

Tempo and pulse — two numbers, kept apart

• pulse_R (circular pulse clarity): onsets folded at the dominant period, strength-weighted resultant length. 0 = free rubato, 1 = metronomic. The period is chosen among the envelope-ACF peak and its metrical octaves by maximising R — because folding at an octave-off period cancels the resultant.
• tempo_bpm (displayed on cards): librosa's perceptually-weighted estimate, targeting the felt beat. The distinction matters: phase concentration genuinely peaks at the tatum (a strummed ballad measured 185 where the felt beat is ~92), so the lock metric and the display tempo must not be the same number.
• Caveats: a single global period conflates slow tempo drift with true rubato; cards prefix the tempo with ~ when pulse_R < 0.1.

Schaeffer proxies (TARTYP & TARSOM)

Object segmentation is onset-bounded; mass comes from spectral-flatness thresholds and centroid drift, facture from duration and 4–20 Hz envelope modulation; the seven TARSOM criteria are signal proxies (see the docs table). Three honest limits:

1. Schaeffer's categories are acts of reduced listening; a flatness threshold is not. The cards are machine-made drafts for the ear, in the spirit (not the practice) of Thoresen's aural sonology.
2. Thresholds are calibrated on tonal instrumental corpora; electro- acoustic material will stress them.
3. The TARTYP grid is the simplified 3×3 of balanced objects; redundant and eccentric classes (homogeneous drones, pedals, large objects) are a planned extension — long drones currently stretch "held N/Y".

Onset handling

librosa's peak-picker normalizes noise into "onsets" on near-featureless signals; all onset-dependent measures apply an absolute onset-strength floor (real attacks measure O(1–10) in dB-difference units, numerical noise O(0.01)).

Prior work

Manual Schaeffer-style visualization has a long tradition (Thoresen's aural sonology and its EAnalysis plug-in, INA-GRM's Acousmographe); automatic TARTYP classification exists as research (Frisk 2025; MTG morphological description; Lartillot's surveys of computational electroacoustic analysis). musiscape's contribution is the assembled, batch-capable pipeline: segmentation → classification → rendered per-track cards for whole collections, as an open tool.