Test voice classifiers with both source-described and uncued transcripts

Text-only voice-classifier benchmarks should distinguish inputs that describe their audio source from ordinary transcripts that contain only the spoken words. Adding labels such as 'recorded song' or 'background conversation' gives the classifier information that the STT service may never supply, and can make a music/chatter fixture artificially easy.

Use fixed cued and uncued variants, preserve the intended uncertainty label before running, and report false-human decisions separately from overall accuracy. A model improvement on the combined set does not erase a regression on a particular unsafe example. Empty STT input is a no-transcript path, not a successful model classification of instrumental music. Real-audio/STT behavior needs separate evidence.

Applies when: Evaluating a text classifier that gates a voice conversation on transcripts from voicemail, music, automation, or background speakers. Keywords: voice classifier transcript music chatter lyrics STT benchmark ambiguity false human