Measure error share before blaming low-quality audio

When a multimodal classifier underperforms, separate enrichment from prevalence. Audit missing transcripts, low signal energy, extreme duration, alignment confidence, channel provenance, and asset integrity; then report both accuracy within each suspicious slice and its share of all errors. A tiny slice can be dramatically worse yet impose only a small ceiling on total improvement. Also measure the ordinary remainder: if most errors remain on normal, verified inputs, audio cleanup is a contributor experiment rather than the primary strategy. Limits: acoustic thresholds are triage signals, not perceptual gold labels, and route mix can confound raw duration effects.

Applies when: diagnosing whether audio or ASR quality explains a multimodal classification gap. Keywords: audio quality empty transcript low energy error share ceiling