Audel — ears for AI agents and media pipelines

Audel hears audio and then listens to it. It returns a machine verdict (pass, warn or fail) with findings grounded to the millisecond, so an agent or a pipeline can fix the problem before it ships.

• Hearing asks whether the signal is healthy and on spec. It grades every track against a delivery spec (EBU R128, R128 s1, ATSC A/85, OP-59, ARIB TR-B32, Netflix-style speech-gated, Spotify, YouTube, Apple), covering integrated, short-term and momentary loudness, true peak, clipping, silence, truncation, dead or inverted or out-of-phase channels, DC offset, mains hum and noise floor.
• Listening asks whether the right thing was said and whether a person could follow it. It checks script adherence with time-grounded omissions and substitutions, required and forbidden wording, pace, dead air, intelligibility, captions (SRT and WebVTT text, offset, reading rate) and language per track. With an opt-in model it also checks whether key messages landed and how they were delivered.

Hearing and offline listening never touch the network. Model-assisted understanding can run entirely on your own hardware through an on-prem Ollama server.

Install

pip install audel                 # light base: ffmpeg + DSP
pip install "audel[asr]"          # local faster-whisper: listening
pip install "audel[serve,oidc]"   # REST service with JWT auth
audel doctor
audel demo                        # broken clip -> FAIL, fixed clip -> PASS, no key

Audel needs ffmpeg on the PATH.

Use it

audel check  promo.wav --spec ebu_r128_s1
audel listen read.wav --script script.txt --captions read.srt --require "terms and conditions apply"
audel qc     master.wav --spec atsc_a85 --html qc.html --json qc.json

from audel import check, listen

heard = await check("promo.wav", spec="ebu_r128_s1")
listened = await listen("read.wav", script=open("script.txt").read())
print(listened.verdict, listened.listening.script_adherence)

Run it as a service

audel serve provides sync grading and async jobs. It can ingest from signed object-store URLs, and it exposes Prometheus metrics and a JSON audit log. Auth is per-tenant tokens or OIDC JWT, with optional mutual TLS, per-tenant rate limits, and an egress policy (allow, local-only or deny). A hardened container and Helm chart ship in the repository. See Running as a service and the threat model.

How accurate is it

audel bench reports precision and recall per check. It runs on a seeded synthetic corpus out of the box, and on your own labelled deliverables with --corpus. Measure it on your material before you make it a gate. See Checks reference.

Part of a family

Audel is the ears. AgentVision is the eyes and Verel is the brain. All three speak one contract, agentsensory, and pip install "verel[hearing]" pulls in the whole stack.

📖 Docs: amitpatole.github.io/audel · Changelog: CHANGELOG.md · Security: SECURITY.md

License

MIT © Amit Patole