Vicidial AMD: the complete guide to answering machine detection

How answering machine detection works inside Vicidial and Asterisk, the amd.conf settings that control it, why stock AMD drops live humans, and how AI AMD fixes it — without changing your dialer or carrier.

What is AMD in Vicidial?

Answering machine detection (AMD) is the step that decides, the instant an outbound call is answered, whether a live human or a machine — voicemail, IVR, or a carrier recording — picked up. In a Vicidial predictive-dialer setup, that decision determines whether the call is handed to an agent or dropped, so AMD accuracy directly drives both your agent efficiency and your abandon rate.

Out of the box, Vicidial leans on Asterisk’s built-in AMD — the app_amd application, configured in amd.conf. It listens to the first few seconds of audio, applies a set of timing rules, and sets an AMDSTATUS channel variable of HUMAN, MACHINE, or NOTSURE. Your dialplan then routes on that value.

How Vicidial’s built-in AMD works

Asterisk’s app_amd is a heuristic, timing-based detector. It measures how long the answerer speaks, how long they pause, and how many “words” it hears, then compares those numbers against thresholds in amd.conf. The main settings:

The logic is intuitive but brittle: a quick “hello,” a short pause, few words ⇒ HUMAN; a long continuous greeting or a beep ⇒ MACHINE. The problem is that real people don’t read the manual.

Why stock Vicidial AMD misclassifies live humans

Timing rules break on the most ordinary human behavior:

• The fast greeting. “Hello? Who’s this?” with no pause is the single most natural way to answer a call from an unknown number — and it’s exactly the pattern app_amd reads as a machine. Why “Hello?” breaks AMD →
• The silent pickup. Someone answers and waits. Timing AMD has nothing to measure and guesses — often wrong.
• Carrier false-answers (FAS). The network signals “connected” when nobody actually picked up. Stock AMD never sees a greeting and mishandles it. Across the AMDY network, about 14% of answered calls are FAS.
• Accents, languages, background noise. Thresholds tuned for one speaking pattern misfire on another.

Heuristic AMD misclassifies roughly 5–15% of live humans as machines. Those are real prospects your dialer hangs up on — and because they’re logged as “answering machine,” they’re invisible in your reports. Read more →

The TCPA timing bind

Tuning isn’t a free fix. To catch more machines, operators make app_amd more aggressive — but that increases false positives (humans called machines), which pushes them to dial harder to keep agents busy, which drives the abandon rate up. The FTC’s Telemarketing Sales Rule expects abandons under 3%. Accurate detection is what lets you dial conservatively and still skip voicemail — staying inside the safe harbor instead of trading compliance for connects.

AI AMD vs stock Asterisk AMD

AI answering machine detection does the same job with a machine-learning model on the answer audio itself instead of hand-tuned timing rules. Because it classifies the acoustic signature of the answer — not a transcript and not a stopwatch — it handles fast greetings, silent pickups, and non-speech answers, in any language.

More on the approach: sound, not words and what AI AMD actually is.

How to set up AMDY on Vicidial

1. Run the installer. One bash command on the Vicidial server provisions an API key and adds the dialplan hook — about 5 minutes.
2. Disable stock AMD for the campaign so AMDY makes the human/machine decision instead of app_amd.
3. Dial a live campaign on the free Sandbox tier and compare your human-connect rate against the old AMD.

Nothing else changes — same carrier, same dialer, same campaigns. AMDY runs through a WebSocket gateway alongside Vicidial, Asterisk, FreeSWITCH, and Issabel.

Frequently asked questions