Findings No. 83
The mono mix drafts AT the stems ceiling: 74/76 = 97.4%
Stem Shovel's mono downmix transcribes at the stems ceiling, 74/76, ahead of the non-drum stems summed
A mono downmix of a track can transcribe about as well as the full set of separated stems. On Stem Shovel, the mono mix drafted 74 of 76 notes against the stems ceiling, 97.4 percent, ahead even of the non-drum stems summed on their own, which only reached 72 of 76.
The cause traces to the drum stem specifically. Basic Pitch, the transcription model Stem Shovel runs, is weak on drums: fed the isolated drum stem, it returns fewer notes than it finds for those same measures in the full mix. Separating the drums did not help Basic Pitch read them better, and the mix still carried enough of the other instruments for the model to transcribe through the bleed between them.
The fix is to stop treating stems as a prerequisite. A stereo upload with no stems attached can go straight into the pipeline as its own mono downmix, rather than waiting on or requiring separation first.
The general rule: before isolating a signal for a model, check whether isolation actually helps that model. If the weakest stage reads a mixed signal as well as or better than the separated one, route around the separation instead of defending it.
Receipt: Stem Shovel, Basic Pitch transcription pipeline, mono mix vs. stems comparison (source document: stem-shovel).