A bass model reads the line far better than Basic Pitch, and exposed notes missing from the truth

Parts

Study Overview

The Reader

The model is the bass checkpoint of an open audio to MIDI package: a network with onset, offset, frame and velocity outputs, of the kind built for piano transcription and retrained per instrument. The code and the model card are MIT licensed. It was run exactly as shipped on Kevin's six bass stems, on a CPU, at about one second of work for three of audio. Its downloaded weights are a pickle file that the library loads unsafely; the bench loaded them through the safe loader with only array types allowed, and they load. A second model, MuScriptor, is registered as another arm and has not run: its weights are non-commercial and sit behind a login that is Kevin's to give.

The Truth Was Missing Notes

On the first pass the new reader scored 43 percent of bars right on Letter Lands, where Basic Pitch scores 76, while beating it note for note. Looking at the bars showed why: the extra notes it was charged with were the same notes, in the same places, that pYIN wrote. Logic writes two notes of one pitch played back to back with the second note's start ahead of the first note's end. The bench's parser paired each end with the latest start, so it lost the first note and left the second with no length, and the grader then dropped that one as too short to be meant. Peaceful Dreams was missing 47 of 454 notes, Interbellum 71 of 491, Downline 77 of 438. Every bass study had been graded against that truth. The parser was fixed and pinned by a test, the alignment of every song was confirmed unchanged, Studies 48 to 51 were re-run and corrected where they stand, and this study was registered again with the numbers from its first pass disclosed.

The Line

On the corrected truth the bass model is the better reader of every song by the note: an octave-exact score of 0.83 to 0.99 against 0.72 to 0.89 for Basic Pitch with its bass filter, and not one octave error in 1,962 matched notes, with no filter at all. It adds a third as many false onsets. Bars entirely right are 53, 39, 51, 82, 74 and 87 percent, a mean of 64 against 48. Three songs, Eat All the Clocks, Downline and Mop Heart, pass the bar for showing the line: at least 70 percent of bars right and at most 10 percent wrong. Where the bass model and Basic Pitch write the same bar, the bar is right 77 percent of the time, over half the bars, and that passes the bar for trusted bars on four songs of six.

Two Songs

Letter Lands is lower than Basic Pitch, 51 percent against 76, and the reason is again a question about the truth. Its MIDI holds twenty notes four milliseconds long, on the fourth beat of the riff's bars. The grader treats a note that short as not meant. The model writes a note in nineteen of those twenty places. With those forgiven it scores 79 percent, above Basic Pitch; that is a what-if, computed after the fact, and whether a four millisecond note is a note is for the composer to say. Interbellum is the other: runs of short notes played without a gap, which every reader here joins or drops. The bass model gets 39 percent of its bars and 54 percent wrong.

Decision

By the rules registered before the run, nothing is recommended. The line is showable on three songs, not on every song. As a replacement for Basic Pitch it gains sixteen points on the mean where ten were required, but one song loses twenty-five where two were allowed. No agreement rule qualifies on every song. Those rules are not moved after the fact. What the study does change is the picture: a model built to follow one note at a time reads a bass stem far better than a chord transcriber with different settings, which four studies of settings, filters, grids and trust rules could not make it do. And it found a fault in the bench's own truth that had been understating every transcriber for five studies.