What the bass line's wrong onsets are: not mostly re-triggers, and no threshold or filter fixes them

Parts

Study Overview

The Wrong Onsets

Of 582 extra onsets across the six songs, 30 percent start while a note of the same pitch class is already sounding, 23 percent start where the part has nothing sounding, 18 percent sit in bars where the truth and the transcription chose different subdivisions, sixteenths against triplets, 13 percent are tails of a note just ended, and 10 percent are a real note one slot off. Of 332 missed onsets, a quarter are in such bars, 30 percent have no transcribed note of their pitch at all, and 17 percent were joined to the note before. The missed notes are the short and the legato ones: on the three songs that have them, a note under 130 milliseconds is missed 48 to 68 percent of the time, a longer one 6 to 25 percent. Corrected 2026-11-12: this paragraph first said the beat tracker put those bars on the other subdivision. The choice is the grader's, made per bar from each line's own onsets, and whether the tracker's beats cause the disagreement has not been measured.

The Transcriber's Own Settings

The app runs Basic Pitch with settings chosen for chords, including a minimum note of about 130 milliseconds, longer than a sixteenth at most of these tempos. From one set of cached model outputs the bench made notes under 32 settings. A lower onset threshold costs 14 to 45 points of bars right on every song. A shorter minimum note finds the short notes and adds more false ones than it finds, worse on every song. A lower frame threshold is worse on every song. The chord settings are right for bass. The only consistent gain is turning off the two passes that make notes without a detected onset: every song improves or holds, by one to seventeen points.

The Filters

Merging a same-pitch note that starts within a sixteenth of the last lifts one song by ten points and drops another by twelve, because that bassist really does repeat sixteenths. Within an eighth it swallows up to a quarter of a song's notes. An amplitude floor lifts the one song with the most stray notes by fifteen points and costs every other song up to seven percent of its real notes. No setting but no filter at all stays inside the registered bound of three percent of real notes lost on every song.

Decision

Nothing is recommended. The selected settings gain 6.2 points on the mean against the ten registered, though no song loses and the choice holds up when each song is held out in turn on five of six. No filter survives. The line stays unshown: 32 to 46 percent of bars right on five songs and 76 on the sixth. The bench's own sentence in Study 48, that the extras were mostly re-triggers and ghost notes, was wrong and is corrected there. What is left is not a threshold: it is the choice of subdivision, and short legato notes the transcriber does not separate. The next studies take those in turn, the grid first, then whether the app can tell which bars to trust. Added 2026-11-12: Study 50 then measured the subdivision and the grid. Choosing the subdivision once per song removes every such bar and bars right does not move, and the composer's own grid scores the same as the tracker's. Neither is where the line is lost.