Dialer Mechanics — the classifier that decides who answered

Answering machine detection: how AMD decides, what it costs, and where it fails

Answering machine detection (AMD) is software that listens to the first seconds of an answered call and classifies the pickup: live human, voicemail greeting, fax tone, or silence. This page covers how both generations of classifier decide, the published parameters, the dead-air and false-positive trade-offs, what AMD costs, and where it helps versus hurts. For prerecorded-message calling itself, see the robo dialer page; for the wider family of machines that place calls from lists, see automated calling systems. Before reading on: Enzo does not use AMD — every connected call goes to a live agent — and the last section explains why that’s deliberate.

What answering machine detection actually is

When an outbound call is answered, somebody — or something — starts making sound. AMD listens to that opening audio and returns a verdict — human, machine, not sure, sometimes fax — and the dialer branches: connect an agent, play a message, wait for the beep, or hang up and dial the next record.

Every documented implementation — Twilio, Telnyx, Plivo, the open-source Asterisk application — runs platform-side, on the calling provider’s media servers, analyzing the answered call’s audio; it is not a service of the phone network itself. Twilio’s docs make the plumbing concrete: its asynchronous AMD consumes one of a call’s four forked audio streams, competing with media streaming, SIPREC, and real-time transcription (Twilio AMD documentation, July 2026).

Generation one: cadence and energy analysis

The classic technique doesn’t understand words — it measures the shape of the audio. Humans answer short and stop: “Hello?” Machines answer long and keep going: “You’ve reached the offices of…” Cadence AMD counts bursts of voiced sound and the silences between them against an energy threshold, and decides from the pattern.

The open-source Asterisk AMD application — the reference implementation behind many dialer stacks — publishes exactly how it decides (Asterisk docs and shipped amd.conf.sample, July 2026). Its defaults: a 5,000 ms total decision window, a greeting cap of 1,500 ms, 800 ms of after-greeting silence to declare a human, a maximum of 2 “words” (voiced bursts), and a silence threshold of 256 on a 0–32767 energy scale. In plain terms: answer with a short hello and pause, and you’re HUMAN; run past a second and a half of greeting or a third burst of speech, and you’re MACHINE before your sentence ends. It also returns NOTSURE and HANGUP — an admission, built into the API, that the pattern is a guess.

Twilio’s AMD is the same family with commercial polish: it “isolates human speech audio and measures periods between speech and silence in the greeting,” trained on real-world call samples per its 2019 launch announcement. Its published defaults draw the same line — utterances shorter than 2,400 ms score human, longer score machine; 1,200 ms of silence marks the end of speech; 5,000 ms of silence returns “unknown”; and the overall detection timeout defaults to 30 seconds (range 3–59).

Generation two: machine-learning classifiers

The newer generation replaces pattern-measuring with trained audio models. Telnyx ships both generations: a standard cadence-style tier, and Premium AMD using “advanced speech recognition technology and machine learning” to return granular classes — human_residence, human_business, machine, silence, fax_detected, not_sure (Telnyx developer docs, July 2026).

Research gives the cleanest accuracy picture available. A 2024 peer-reviewed MIPRO paper (arXiv 2410.08235) using YAMNet transfer learning with an RNN classifier on streaming audio reports over 96% accuracy on its test set, exceeding 98% when combined with a silence-detection stage.

Those research figures are the only clean accuracy numbers in this space. No major platform publishes false-positive or false-negative rates: Twilio and Plivo publish no accuracy figure at all, and Telnyx’s own explainer says only “exceptional accuracy.” Percentages circulating in dialer marketing are unaudited claims, not measurements — and the research numbers were earned on the researchers’ test set, not on your list.

The four documented implementations, side by side

Beep detection and the voicemail-drop pipeline

AMD’s second job is timing: message-delivery campaigns need to know when the greeting ends, so a recording lands after the beep instead of over it. Twilio splits its AMD into two modes accordingly: “Enable” returns a verdict as soon as one is reached, while “DetectMessageEnd” holds machine verdicts until the end of the greeting, “usually indicated by a beep,” returning machine_end_beep, machine_end_silence, or machine_end_other. Telnyx does the equivalent with beep_detected / no_beep_detected webhooks.

Plivo shows a third shape: asynchronous only. It analyzes for 5,000 ms by default and delivers the verdict by HTTP callback while the call keeps going — the flow is never blocked, so the application transfers or hangs up after the fact. That suits broadcast and message flows, and is unusable as a gate in front of a live agent.

This machinery is what “voicemail drop” products are built on; Enzo has neither AMD nor voicemail drop.

The trade-offs, in the vendors’ own words

Dead air. A classifier needs audio before it can decide, and the person who answered says “Hello? … Hello?” into that pause. Nobody publishes a mean-latency figure, but the defaults bound the problem: Plivo allots 5,000 ms of analysis, Asterisk’s decision window is 5,000 ms, and Twilio’s and Telnyx’s detection timeouts default to 30 seconds. Twilio states the mechanism outright: raising the speech-end threshold to fix a false positive means “increasing the delay for human detection by the amount you increase this parameter.” Faster verdicts or fewer errors — the dial doesn’t turn both ways at once.

False positives on live humans. Twilio’s docs walk through both directions of error: a “very short two-second voicemail greeting” can be misread as a human picking up, and a human whose answer comes as “two utterances separated by a period of silence” may be interpreted as a machine. The second error is the expensive one — a real person answered, and the system dropped them or greeted them with silence. Accuracy is therefore greeting-dependent: the same classifier scores differently against different lists — one more reason a single marketing percentage tells you little.

The asymmetry. A false negative — voicemail routed to an agent — wastes a few seconds. A false positive burns a live answer and (next section) counts against a regulatory cap, which is why sensible tuning biases toward fast, conservative human verdicts, accepting more voicemails-to-agents as the price.

AMD and the 3% abandonment standard

Under the FTC’s Telemarketing Sales Rule, an outbound telemarketing call answered by a person counts as abandoned unless a representative is connected within two seconds of the person’s completed greeting, and the rule’s safe harbor caps abandoned calls at 3% of person-answered calls per campaign or 30-day period (16 CFR § 310.4) — the full mechanics are on our Telemarketing Sales Rule guide.

The interplay follows from the arithmetic. Machine-answered calls are excluded from the abandonment denominator, so AMD errors on machines are regulatorily free. But a live person misread as a machine and dropped — or held in silence past the two-second mark while the classifier decides — is structurally a call a person answered that never reached a rep: exactly the silent hang-up the rule targets. That is a reading of the rule text alongside vendor defaults, not an FTC pronouncement — if you run AMD at volume, have a compliance attorney review how your abandonment is measured.

Where AMD helps — and where it hurts

AMD earns its keep in message-delivery operations. Consented notification and reminder campaigns, voicemail-drop workflows, and broadcast use cases are machine-vs-human triage at scale: the job is to land a recording cleanly or route the rare live pickup, and beep-timing modes exist precisely for that. The robo dialer page linked above covers that world, including its much stricter consent rules.

AMD costs you in live-conversation selling. When the campaign’s point is a human conversation, the person who answers is the scarcest asset in the funnel — and AMD’s failure modes spend exactly that asset: the detection pause greets your best outcome with silence, and a false positive hangs up on them. A voicemail slipping through to an agent costs seconds. For teams selling by voice, the classifier defends against the cheap error and risks the expensive one.

The market has noticed, and rebranded. In the AI parallel-dialer segment, the same capability is sold as “AI answer detection” or “human answer detection”: ML classifiers screen several simultaneous lines for voicemail greetings, IVR trees, fax tones, and busy signals, connecting the rep only on a live pickup — vendors typically claim connection within 1–3 seconds of the answer (vendor marketing, 2026; no independent audit). Same classifier, inverted question. How that segment compares to plain power dialing is covered in AI parallel dialer vs power dialer.

What AMD costs on the major platforms

Published prices are small per call and real at volume. Twilio charges $0.0075 per call where AMD is enabled and the called party answers — busy and failed calls aren’t charged — on top of standard voice rates (2019 launch announcement; still the documented rate as of July 2026); at 100,000 answered calls a month, that line item is $750. Telnyx’s December 2022 release note set standard AMD at $0.0020 and Premium at $0.0065 per invocation — figures old enough to verify against current pricing before budgeting. Plivo publishes no price. No platform sells an accuracy commitment with the fee; Telnyx’s own guidance is to watch your misclassification counts yourself.

Enzo doesn’t use AMD — on purpose

Enzo doesn’t use AMD; every connect goes to a live agent, so there’s no detection delay to hang up on. In every Enzo mode — power, predictive, preview, single-line, multi-line — the voice an answerer hears is your agent’s, because no classifier sits between the pickup and the person. The trade is honest: agents hear the occasional voicemail greeting and disposition it, in exchange for never greeting a live answer with algorithmic silence. There’s no prerecorded voice and no AI voice anywhere in the product.

That’s a fit statement, not a verdict on the technology. AMD is well-built machinery for message-delivery campaigns, and the platforms above sell it openly — examples, not endorsements. If voicemail drops and broadcast triage are your workflow, Enzo is the wrong tool and we’ll say so. If you sell by live conversation: pricing is published at /pricing — from $99 per seat per month billed annually, $120 month-to-month, no seat minimum, all inbound and outbound minutes included — and the evaluation path is a free 20-minute discovery call. Book a discovery call — if Enzo isn’t the right fit, we’ll tell you.

Sources: Asterisk AMD documentation and shipped amd.conf.sample defaults; Twilio AMD docs and 2019 general-availability announcement; Telnyx AMD developer docs and December 2022 pricing release note; Plivo machine-detection docs; Altwlkany et al., 47th MIPRO 2024 (arXiv:2410.08235); 16 CFR § 310.4 — all as of July 2026. Vendor parameters, prices, and behavior change without notice; verify with each vendor before you build on them.