Audel — Ears for AI Agents¶

Problem: AI coding agents are deaf — they generate TTS, soundtracks, web-app sound effects, voice-bot prompts, and audiobooks, and never hear the result. Result: Audel gives them ears — play → hear → report → fix — so the agent self-corrects before it claims done.

pip install audel
audel demo                       # no API key required

audel demo grades a deliberately silent render, prints a FAIL report (missing audio — DSP-grounded, no LLM key), then loops against the fixed version and prints "what changed: issues resolved → PASS."

What it does¶

Capability	What you get
Hear & report	A machine verdict (`pass`/`warn`/`fail`) + time-grounded issues — loudness (EBU R128), clipping, silence/dropouts, A/V desync — DSP-grounded, no key needed.
Match intent	Grade audio against a brief — language, phrase coverage, duration — via local faster-whisper ASR. PASS means "it says what I meant," not just "it has sound."
Listen & critique	`analyze` adds acoustic + LLM judgment (Ollama Cloud text critique, or key-gated Anthropic / Gemini-audio / CLAP) for subjective claims — egress only here; `check` stays offline.
Streaming / temporal	`watch` grades behavior over time; `StreamMonitor` grades live PCM chunks — playthrough, dropouts, liveness — bounded by construction.
Ears → brain handoff	A distilled `{verdict, next_action, todo}` signal any agent/brain acts on.

Where to go next¶

:material-rocket-launch: Quickstart — install, system deps, first run.
:material-console: CLI — demo · check · analyze · watch · render · diff · stream · serve · doctor.
:material-cog: Adapters — REST (FastAPI) and MCP (9 tools) service surfaces.

Offline by construction¶

check never touches the network — loudness, clipping, silence, dropout and A/V-desync grading run entirely on local DSP (ffmpeg filters) and local ASR (faster-whisper). Egress happens only on analyze, and only to a backend you name; transcripts are treated as untrusted input. The base install is light (ffmpeg + DSP, no torch) — ML lives behind extras: [asr], [render], [cloud], [clap].

Eyes, ears & brain¶

Audel is the ears. It pairs with AgentVision, the eyes — which confirms a video renders — and Verel, the brain — an agent framework where nothing is "done" until a grader returns a verdict. AgentVision confirms it renders; Audel confirms it actually has sound and the narration is right; Verel decides and compounds only verified work into memory. All three share one contract, agentsensory.

  AgentVision  ·  Audel  ·  Verel
     eyes          ears       brain
   it renders   it sounds   it's done

Install: pip install "verel[hearing]" pulls the whole stack · Source: GitHub · Package: PyPI · License: MIT.