Abstract

Grid-guided onset detection scores a musician’s timing by asking, at every position of a metronome grid, whether a note began there. On sustained-tone instruments (brass, single-reed, flute) this fails in one specific way: a note held across several grid positions keeps producing spectral change through vibrato, breath and pitch wobble, so its interior is reported as a run of new notes. Every test that reads the sound at the position (spectral flux, prominence, level, spectral sparsity, group delay, linear-prediction residual, pitch) is either inert on this population or removes only a minority of it at a cost in real notes no better than simply raising the detection threshold, because a held note genuinely is sounding at full level there. This disclosure describes a candidacy test asked before any onset question: was the previous note ever interrupted? A grid position is excluded from candidacy, placed in a third state beside detected and missed and removed from every denominator, when the short-window loudness envelope in a window around it never fell below a fixed fraction of its lower shoulder and the settled pitch either side of it moved by less than a threshold. The test fails open on every unmeasurable input, so an entry from silence is always examined, and it can only remove candidates, never add them. Variants of the window, ratio, envelope resolution, shoulder definition, edge-trough reading, pitch measurement and denominator treatment are disclosed, together with two planned companion rules. On ten labelled free-practice recordings (trumpet, alto saxophone) it removed 35 of 49 held-note false positives at the cost of one marginal note; on 24 protocol recordings across six wind instruments it removed 44 of 46 with no note lost.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS