amd

Answering-machine detection for Fonoster. An ONNX whisper-tiny classifier reads the first seconds of a call's audio and returns a status (HUMAN, MACHINE, VOICEMAIL, IVR or UNKNOWN), a confidence, the detector name and the latency.

AppBridge loads it for the amd listener: it attaches a listen-only participant to the room, collects the caller's audio (16 kHz L16) and calls classifyPcm. The model ships with the package in amd-model/.

import { classifyPcm } from "@fonoster/amd";

const verdict = await classifyPcm(pcm);

Regression corpus

test/fixtures/regression/ is a permanently-stored, correctly-licensed corpus of 21 audio clips (each 5s or longer) covering all four classifier labels — 11 human (10 real conversational clips from PolyAI/MINDS-14, CC BY 4.0, plus one scripted-narration edge case), 3 voicemail and 1 answering- machine clip (real, CC0, from Freesound), and 2 answering-machine + 4 IVR clips synthesized in-house (no real-world recording with a clean license turned up for those two categories). manifest.json pins each clip's current classifier output (knownStatus/knownConfidence), which test/amd/regression.test.ts asserts against on every run — the goal is catching a silent behavior change (model swap, feature-extractor edit, dependency bump) rather than measuring accuracy. Where the pinned baseline disagrees with the file's real-world label (groundTruthLabel), that's a documented known accuracy gap, not a test bug — see each entry's note; a few exist on purpose, to track known confusions (e.g. formal/synthesized IVR scripts reading as MACHINE) rather than hide them. Licensing and attribution for every third-party clip is in test/fixtures/regression/ATTRIBUTION.md.