A compact boundary model can recover useful note events from CREPE pitch contours in small-data monophonic humming when annotations are audited and parameter selection is nested, according to a methodology study published in the Journal on Audio, Speech, and Music Processing. The investigation demonstrated that pairing a full CREPE pitch estimator with targeted classifiers achieves promising segmentation within a controlled corpus, although significant performance declines on external data indicate that corpus-independent generalisation has not yet been established.
Humming provides a convenient input format for query-by-humming systems and music-education tools, but establishing precise note boundaries remains inherently difficult. Amateur humming performances frequently introduce weak boundaries characterized by glides, vibrato, intonation error, and short unvoiced gaps. To address these acoustic challenges, the researchers combined the full CREPE pitch estimator with 44 frame-level features capturing pitch, energy, spectral flux, and voicing context.
For the segmentation stage, the researchers deployed separate RandomForest classifiers designed to estimate onset and offset probabilities independently. To prevent data leakage and avoid optimistic bias during evaluation, model and decoder parameters were tuned strictly within a recording-grouped five-by-four nested cross-validation protocol. Under this design, each outer test recording was systematically excluded from every parameter selection step.
The empirical validation relied on 38 MTG-QBH/Molina recordings containing 2,152 reference notes following an audit of the annotations. Within the primary nested cross-validation evaluation, the proposed framework achieved micro-averaged COnPOff, COnP, and COn F-measures of 0.696, 0.842, and 0.889, respectively. The corresponding macro-averaged COnPOff F-measure reached 0.685, with a 95 percent recording-bootstrap confidence interval spanning from 0.643 to 0.727. In terms of frame-level performance, the system attained a raw pitch accuracy of 0.882 and an overall accuracy of 0.761.
The researchers compared their approach directly against an existing benchmark using the same 38 waveforms and evaluation metrics. A fixed official ROSVOT checkpoint, originally trained on the M4Singer dataset, obtained COnPOff, COnP, and COn scores of 0.171, 0.230, and 0.265, respectively, when evaluated without Molina tuning. Furthermore, an analysis of segmentation events revealed that detections were not merely clustered around phrase boundaries. Only 2.4 percent of segmentation events occurred within 150 milliseconds of an operational phrase boundary, substantially lower than the 14.5 percent that would be expected based solely on temporal coverage.
To assess generalisation beyond the primary dataset, the team conducted external evaluations on the HumTrans corpus. In a native zero-shot HumTrans evaluation, the system obtained a COnPOff score of 0.008. When a separate validation-aligned sensitivity analysis was applied using a frozen negative 0.18-second shift, the resulting COnPOff score reached 0.146. Additionally, the study evaluated tonal representations by examining audio-derived and note-derived pitch-class profiles, which exhibited high descriptive agreement with a mean cosine similarity of 0.978. However, the authors emphasized that this descriptive agreement must not be interpreted as key-prediction accuracy, as independent key labels were unavailable in the corpus.
The study concluded that while compact boundary modeling can extract note events from audited humming recordings, the substantial internal-to-external performance drop demonstrates that the current evidence cannot establish corpus-independent generalisation. The authors emphasized that the 38-recording annotation set remains the limiting unit of independent evidence in this research. Furthermore, remaining transcription errors occurred primarily within continuously voiced passages, prompting the researchers to recommend that tonal descriptors remain strictly separate from hard pitch correction mechanisms.
Subject of Research: Note-level melody estimation and boundary segmentation in monophonic humming
Article Title: CREPE-guided note-level melody estimation and tonal-representation analysis for monophonic humming
Article References: Liu, J., Wang, C., & Jiang, H. (2026). CREPE-guided note-level melody estimation and tonal-representation analysis for monophonic humming. Journal on Audio, Speech, and Music Processing. https://doi.org/10.1186/s13636-026-00482-1
Image Credits: AI Generated
DOI: 10.1186/s13636-026-00482-1
Keywords: monophonic humming, CREPE pitch estimator, note-level melody estimation, RandomForest classifiers, nested cross-validation, onset detection, offset detection, MTG-QBH dataset, Molina recordings, M4Singer dataset, ROSVOT checkpoint, COnPOff metric, raw pitch accuracy, HumTrans test, pitch-class profiles, phrase boundaries, spectral-flux features, voicing-context features
Cite Scienmag News
Courtney Benton. (November 17, 2026). Compact Model Recovers Humming Notes but Shows External Performance Drop. Scienmag. https://scienmag.com/compact-model-recovers-humming-notes-but-shows-external-performance-drop/
Courtney Benton. "Compact Model Recovers Humming Notes but Shows External Performance Drop." Scienmag, 17 November 2026, https://scienmag.com/compact-model-recovers-humming-notes-but-shows-external-performance-drop/. Accessed 17 November 2026.
Courtney Benton. "Compact Model Recovers Humming Notes but Shows External Performance Drop." Scienmag. November 17, 2026. https://scienmag.com/compact-model-recovers-humming-notes-but-shows-external-performance-drop/








