Two rules on the tracker's bars, measured: joining is a wash with one clear win, and the steady share enriches without separating

Listening

Study Overview

The Rules, Ported

The app's grid code was ported line for line and pinned to its own thirteen test cases, so the bench measured the rule the app ships and not a reading of it. All five of the app's dev recordings were already in the bench's corpus with package readings committed, and the app's beats-per-bar table matched those readings count for count on all five, which answers the port comparison at the bar level before any new work.

Joining

Over the census, joining moves the tracker's bar count toward the declared count on 73 songs and away on 61; among the 180 songs the tracker reads at the declared level it is 44 to 36, with the mean count error falling from 8.7 to 6.1 percent. Two of the app's own five moved toward and two away, the latter two having been under the declared count before joining, so their joined pairs were by count real bars. The clear win is a 6/8 song the tracker reads as bars of three and six by turns: joining turns it into 94 bars against 94 declared and lifts its steady share from 0.46 to 0.84, without which a good draft fails the second rule. The bench's reading is to keep joining, because its failure is cheaper than its success is valuable.

The Steady Share

At the app's threshold the share removes 35 census readings, 13 with the bar wrong at every fold and 22 with it right at some fold, so it enriches for a wrong bar without separating. Among the 22 drafts the bench holds on this tracker's bars it removes two sound drafts and misses the one confidently wrong draft, whose share is 0.93. A fold-tolerant share and the beat interval's coefficient of variation were swept beside it and did no better at any threshold. The registered criterion lands on 0.7, which the bench reports and does not like; its labelled judgment is 0.4, the least that keeps the one chaotic reading off the chart while keeping both good 6/8 drafts.

The Fast Waltz

The 3/4 at 224 beats per minute is not a weakness of the small checkpoint: the big one fails the same rule with the same beat and a different modal bar, and only the dynamic Bayesian postprocessor, like madmom, reads it steadily, at half the declared tempo. The program's other 3/4 stands under both checkpoints, the census's one 3/4 is steady, and its seventeen fastest songs have a median share of 0.88. Fast is not the problem; this song's bar is. The fold-tolerant share tells the case apart from the small checkpoint's output alone on this corpus; the beat interval's variation does not.