Skip to main content
Lightning Jar - Web Studio Lightning Jar Wordmark

Findings No. 85

Do not HPSS

HPSS scores near perfect on clean synthetics, then drops bars 72 and 76 in Stem Shovel when it reads strums and vibrato as percussive

Stem Shovel Browser Platform verified

Harmonic percussive source separation looks like a natural preprocessing step before chord detection, but it should not be trusted as a default filter ahead of one. Stem Shovel's study S33, test G3, failed when HPSS was added to the pipeline: the harmonic component lost two bars, 72 and 76, because the median filter used to split harmonic from percussive content read strummed attacks and vibrato as percussive and discarded them. What made the failure easy to miss is that the same separation looked flawless beforehand. On clean synthetic test material it produced 99.6 to 99.7 percent chord-tone purity at every pan width tested, which is exactly the trap: clean synthetics don't carry the attack transients and vibrato that real guitar and keyboard parts do, so a filter tuned against them can pass every check and still fail on the material it is meant for. The fix was to publish the failing run as found rather than quietly patch around it, so S33-G3 stands as a documented case against reaching for HPSS as a chord-detection preprocessor. The general rule: a filter that scores near perfect on clean synthetic data has not been tested against the signal characteristics, attacks, vibrato, that break it in real recordings, and a near-perfect synthetic score is a reason to look harder, not a reason to stop.

Receipt: Stem Shovel, study S33 test G3 (bars 72 and 76).

Ask Eljay

Ask about the studio's work, research, packages, or writing. A few starters:

Answers come from this site's own content and link their sources. For anything that matters, email hello@lightningjar.com.