audit-audio

audit-audio — audio-only quality gate

Fresh-context agent. Receives ONLY the audio track (video stripped). Cannot be biased by seeing the visuals. Tests what a LISTENER would hear.

How to run

python3 ${CLAUDE_PLUGIN_ROOT}/skills/audit-audio/scripts/check.py \
    --clip 20_DELIVER/v1/clip.mp4 \
    --out 10_WORK/audit_audio.json

What it checks

1. WORD CLIPPING — any words cut off?

• Transcribe the audio independently; ASR word durations are HYPOTHESES, not verdicts
• The forced aligner glitches at clip start: the FIRST word routinely reports a tiny duration on clips whose waveform attack is clean — first-word ASR flags are suppressed (reported as suppressed_aligner_glitches), never failed
• Zero-gap opens / hard-ends legitimately put near-full voice level at the literal clip edges, so a hot edge alone is not evidence either
• FAIL only when the LAST word is ASR-short AND the waveform confirms the audio ends at full voice level (cut mid-decay) — corroboration on both channels
• Head/tail edge levels are always reported (head/tail_edge_db_below_speech) for the reviewing agent to eyeball

2. POPS AND CLICKS — any splice artifacts?

• At every edit point (detected via waveform discontinuity), check for:
  • Click/pop: sudden amplitude spike >6dB above surrounding 50ms
  • DC offset jump at splice boundary
• Extract the waveform around each detected splice, measure peak-to-surrounding ratio
• FAIL if any pop/click detected

3. LEVEL CONSISTENCY — voice volume steady throughout?

• Measure LUFS per segment (between splice points)
• All speech segments should be within ±3 LUFS of each other
• No segment should be >6 LUFS different from the average
• FAIL if any segment is >6 LUFS off, WARN if >3 LUFS off

4. MUSIC BED BALANCE — music sitting correctly under voice?

• The declared music bed is read from the render metadata (clip.contract.json / manifest.json via _shared/clip_meta.py); no music declared → check passes with note
• Bed level is measured in isolatable speech gaps (≥0.4s), compared against voiced level
• Target separation is 10–13dB; FAIL if the bed is within 6dB of voice, WARN within 9dB. No gap long enough to isolate the bed → pass with note (ear check)