Separate raw call volume from verified training pairs

Assess voice fine-tuning readiness with separate counts for raw calls, intended-speaker spans, text-aligned candidates, and human-verified audio/text pairs. Mixed mono calls can contain salvageable spans, so missing separate tracks does not establish zero candidate material. Conversely, exact ASR/text agreement does not establish truth: repeated phrases, the wrong speaker, or incorrect boundaries can pass. Track short-reply exclusions explicitly; a strict five-token screen can omit much of a workflow's real speech. Use this funnel to scope curation, not to claim transcription accuracy or sufficient training hours.

Evidence/example — 2026-09-22: a completed Check Calls audit decoded and transcribed 3,708 recordings (46.69 whole-call hours). The filtered driver cohort contained 3,342 stored turns, but 2,041 were below the three-token comparison cutoff. Strict interval/dedup/opposite-role filtering yielded 215 driver candidates and 2,112 current-voice assistant candidates spanning 552 distinct normalized texts. These are machine-screened candidates; no human listening, corrected gold labels, or fine-tuning improvement was established. The practical inference is to curate the existing corpus before concluding that more generic calls are needed, while retaining clean original bot audio for higher-fidelity TTS work.

Source: https://github.com/e3-solutions/check_calls/blob/5587a5fee/docs/voice_data_readiness_audit_20260922.md Implementation/review: https://github.com/e3-solutions/check_calls/pull/166