CREPE-guided note-level melody estimation and tonal-representation analysis for monophonic humming

Abstract

Humming is a convenient input for query-by-humming and music-education systems, but note boundaries are often weak because amateur performances contain glides, vibrato, short unvoiced gaps and intonation error. We combine the full CREPE pitch estimator with 44 frame-level pitch, energy, spectral-flux and voicing-context features. Separate RandomForest classifiers estimate onset and offset probabilities. Model and decoder parameters are selected within a recording-grouped 5 \(\times\) 4 nested cross-validation protocol, so each outer test recording is excluded from every selection step. After auditing the annotations, 38 MTG-QBH/Molina recordings contained 2152 reference notes. The primary nested-CV estimate achieved micro-averaged COnPOff, COnP and COn F -measures of 0.696, 0.842 and 0.889, respectively; the macro COnPOff was 0.685 (95% recording-bootstrap interval, 0.643–0.727). On the same 38 waveforms and metrics, the fixed official M4Singer-trained ROSVOT checkpoint obtained 0.171, 0.230 and 0.265 without Molina tuning. Frame-level raw pitch accuracy was 0.882 and overall accuracy was 0.761. Segmentation events were not concentrated at phrase boundaries: only 2.4% occurred within 150 ms of an operational phrase boundary, compared with 14.5% expected from temporal coverage. In a native zero-shot HumTrans test, COnPOff was 0.008; a separate validation-aligned sensitivity analysis using a frozen –0.18 s shift yielded 0.146. Audio- and note-derived pitch-class profiles showed high descriptive agreement (mean cosine similarity 0.978), but this is not interpreted as key-prediction accuracy because independent key labels are unavailable. A compact boundary model can recover useful note events from CREPE contours in small-data monophonic humming, provided that parameter selection is nested and annotations are audited. The large internal-to-external performance drop shows that this evidence does not establish corpus-independent generalisation. The 38-recording annotation set remains the limiting unit of independent evidence. Remaining errors occur mainly within continuously voiced passages, and tonal descriptors should remain separate from hard pitch correction.

The authors' abstract, as published at the source. EURASIP Journal on Audio Speech and Music Processing, 2026 · DOI ↗