A machine verdict (pass/warn/fail) with time-grounded issues — loudness (EBU R128), clipping, silence, dropouts, A/V desync. DSP-grounded, no key needed.
Grade what the audio says against a brief — language, required phrases, duration — via local faster-whisper ASR.
analyze adds acoustic + LLM judgment for subjective claims. Egress happens only here; check stays offline by construction.
watch grades behaviour over time; StreamMonitor grades live PCM chunks — playthrough, dropouts, liveness — bounded by construction.
audel demo grades a deliberately silent render, prints a FAIL report (missing audio — no LLM key), then loops against the fixed version and prints "what changed: issues resolved → PASS."
It renders — visual feedback, grounded.
It sounds — audio feedback, time-grounded.
It's done — decide, compound verified work.
pip install "verel[hearing]" pulls the whole stack. All three share one contract, agentsensory.