Answering machine detection (AMD) is the part of an outbound dialer that listens to the first second or two of an answered call and decides whether a human or a machine picked up. If it hears a person, the auto dialer connects a sales rep. If it hears a voicemail greeting, the dialer hangs up or leaves a recorded message.

It sounds simple, but AMD is a real-time audio classification problem with a hard deadline. This post explains where AMD sits in a dialer architecture, how classic and AI-based detection work, and the trade-offs you have to design around. The code samples are in Python.

Where Answering Machine Detection Sits in an Outbound Dialer Architecture

AMD runs between "the call was answered" and "the call was routed". In a typical cloud dialer, an outbound call moves through these stages:

  1. Dial. The dialer sends a SIP INVITE to the carrier.
  2. Call progress. The carrier returns ringing (180) or early media (183). A busy line or a bad number ends the call here.
  3. Answer. A 200 OK means something picked up. Audio (RTP) starts to flow.
  4. Detect. The AMD engine analyzes the first 1 to 4 seconds of inbound audio.
  5. Classify. The result is usually human, machine or unknown.
  6. Route. The dialer bridges a rep, drops a voicemail, or disconnects and logs the outcome.

Call progress analysis vs answering machine detection

The two are often mixed up. Call progress analysis (CPA) happens before the answer and covers signals like busy tones, ringback and the three-tone "number not in service" message. Answering machine detection happens after the answer and only asks one question: person or machine?

Why AMD Matters in a Power Dialer, Parallel Dialer and AI Dialer

Most outbound calls are not answered by a person. Without AMD, sales reps spend much of their day listening to voicemail greetings.

AMD matters most in a parallel dialer, which calls several lines per rep and connects the first live answer. The dialer needs to know which of those lines reached a human, and it needs to know quickly. In a power dialer (one line per rep), AMD still saves time by skipping machines and triggering automatic voicemail drop.

How Classic AMD Works: Cadence and Timing Heuristics

Classic AMD does not understand words. It measures the rhythm of sound and silence, because people and machines answer the phone differently.

  • A person says something short, like "Hello?", then waits.
  • A voicemail greeting talks for several seconds without a pause.

The detector splits audio into small frames, marks each frame as speech or silence, and applies timing rules. This is how the AMD application in open-source PBX software such as Asterisk works.

import numpy as np FRAME_MS = 20 SAMPLE_RATE = 8000 # narrowband telephony audio FRAME_LEN = SAMPLE_RATE * FRAME_MS // 1000 def frame_is_speech(frame: np.ndarray, threshold: float = 500.0) -> bool: """Energy-based voice activity detection on one 20 ms PCM16 frame.""" rms = np.sqrt(np.mean(frame.astype(np.float64) ** 2)) return rms > threshold def classify_cadence( pcm: np.ndarray, max_initial_silence_ms: int = 2500, max_greeting_ms: int = 1500, after_greeting_silence_ms: int = 800, max_analysis_ms: int = 4000, ) -> str: initial_silence = speech = trailing_silence = 0 heard_speech = False n_frames = min(len(pcm) // FRAME_LEN, max_analysis_ms // FRAME_MS) for i in range(n_frames): frame = pcm[i * FRAME_LEN:(i + 1) * FRAME_LEN] if frame_is_speech(frame): heard_speech = True speech += FRAME_MS trailing_silence = 0 if speech > max_greeting_ms: return "machine" # long, unbroken greeting elif heard_speech: trailing_silence += FRAME_MS if trailing_silence >= after_greeting_silence_ms: return "human" # short greeting, then waiting else: initial_silence += FRAME_MS if initial_silence > max_initial_silence_ms: return "machine" # long silence before any audio return "unknown" 

The thresholds above are illustrative. Real systems tune them per carrier and per country.

Where cadence-based AMD fails

  • A person who answers with a long sentence ("Thanks for calling Acme, this is Dana speaking") looks like a machine.
  • A short voicemail greeting ("Hi, leave a message") looks like a person.
  • Background noise, music and poor audio break simple energy thresholds.
  • Call-screening assistants on modern smartphones do not fit either pattern.

Voicemail Beep Detection With the Goertzel Algorithm

To leave a voicemail, the dialer has to wait for the beep. A beep is a short, steady tone, so you do not need a full FFT to find it. The Goertzel algorithm measures the energy at one target frequency cheaply, frame by frame.

import math def goertzel_power(frame, target_hz: float, sample_rate: int = 8000) -> float: n = len(frame) k = round(n * target_hz / sample_rate) coeff = 2 * math.cos(2 * math.pi * k / n) s_prev = s_prev2 = 0.0 for x in frame: s = x + coeff * s_prev - s_prev2 s_prev2, s_prev = s_prev, s return s_prev2 ** 2 + s_prev ** 2 - coeff * s_prev * s_prev2 

Beep frequencies vary by carrier and voicemail system, so production detectors scan a set of candidate frequencies. They look for one frequency that dominates the frame for a few hundred milliseconds, followed by silence.

How AI Answering Machine Detection Works

AI answering machine detection replaces hand-written timing rules with a trained audio classifier. The pipeline has four steps:

  1. Feature extraction. Convert the audio stream into log-mel spectrograms or MFCCs over short windows.
  2. Model. Run a small neural network (commonly a CNN or a recurrent model) that outputs a probability for each class.
  3. Streaming decision. Re-score every 100 to 200 ms as more audio arrives, and stop as soon as confidence crosses a threshold.
  4. Fallback. If confidence stays low when time runs out, return unknown and apply a safe default.

An AI dialer learns signals that rules miss: the compressed sound of a recorded greeting, the lack of natural background noise, the pacing of a carrier's default voicemail message. It is also retrainable when new patterns appear.

The model has to be small enough to run in real time across hundreds of concurrent calls, so inference cost per stream is a design constraint, not an afterthought.

The Hard Trade-Off: AMD Speed vs Accuracy

Every AMD engine trades latency for accuracy. More audio means a better decision, but the caller is waiting.

  • False positive (human classified as machine). The dialer hangs up on a real prospect. You lose the lead, and it counts as an abandoned call.
  • False negative (machine classified as human). A rep is connected to a voicemail greeting and wastes a few seconds.

False positives are the expensive error, so most systems are tuned to send uncertain calls to a rep.

The two-second rule

For U.S. telemarketing calls, FCC rules treat a call as abandoned if a live rep is not connected within two seconds of the person's completed greeting. Abandoned calls must stay at or below 3% of live-answered calls per campaign over 30 days. That gives AMD a practical latency budget: detection plus bridging has to fit inside two seconds. This is general information, not legal advice.

Synchronous vs asynchronous AMD

  • Synchronous AMD holds the call until the classifier decides. It is simple to build, but the prospect hears dead air and may hang up.
  • Asynchronous AMD connects the rep right away and keeps classifying in the background. If a machine is detected, the dialer pulls the rep out and drops a voicemail.

Asynchronous AMD removes the awkward pause, at the cost of more complex call-state handling.

How to Measure AMD Accuracy in an Auto Dialer

A single "accuracy" number hides the errors that matter. Track these instead:

  • Human recall: of all real people, how many were classified as human?
  • Machine precision: of all calls classified as machine, how many really were?
  • Detection latency: p50 and p95 time from answer to decision.
  • Unknown rate: how often the engine could not decide.
  • Abandon rate: calls where a person answered and no rep was connected in time.

Measure on your own call traffic. Results change with carrier mix, country, mobile versus landline share, and time of day.

Answering Machine Detection in BelSmart's Dialer Architecture

BelSmart is a cloud-based outbound dialer that runs AI answering machine detection across its dialing modes. Its dialer architecture page describes how AMD fits into the stack:

  • AI AMD in under 600 ms. The engine analyzes audio patterns, not only silence length, to separate live answers from voicemail greetings.
  • Automatic routing. Live answers go to a rep, and voicemails receive an automated voicemail drop.
  • Carrier infrastructure. Calls run over Twilio's elastic SIP trunking, so there is no SIP hardware or PBX to manage.
  • WebRTC in the browser. Reps call from Chrome or any modern browser with no softphone to install.
  • Multi-region failover. Traffic reroutes to another region in under two seconds if a carrier node degrades.
  • Encryption on every call. TLS for SIP signaling and SRTP for media, with recordings encrypted at rest.
  • 99.99% uptime SLA.

For the full call flow, from the rep's browser click to a live prospect, see BelSmart's dialer architecture under the hood.

Try the BelSmart auto dialer free for 7 days

Want to see AI answering machine detection on real calls instead of building it yourself? Sign up in minutes, upload a lead list and start dialing. No credit card required.

Start your 7-day free trial →

Answering Machine Detection FAQ

What is answering machine detection in an outbound dialer?

It is a feature that analyzes the first moments of an answered call to decide whether a person or a voicemail system picked up, so reps only spend time on live conversations.

How accurate is answering machine detection?

It depends on the method and the call traffic. Cadence-based AMD makes more mistakes on short greetings and noisy lines. AI-based AMD is generally more accurate, but no system is perfect, so measure it on your own calls.

How long does AMD take?

Typically between one and four seconds. Faster detection reduces dead air, and for U.S. telemarketing the whole process should fit within two seconds of the person's greeting.

What is the difference between AMD and voicemail drop?

AMD detects that a machine answered. Voicemail drop is what the dialer does next: it plays a recorded message after the beep.

Conclusion

Answering machine detection is a small component with a large effect on an outbound dialer. Classic AMD uses cadence rules, AI AMD uses a streaming audio classifier, and both are limited by the same deadline: decide in about two seconds without hanging up on real people. If you are designing a dialer, treat AMD as a latency-critical service, tune it for human recall first, and measure it on your own traffic.

Try BelSmart free for 7 days →