Answering Machine Detection: How AMD Works, Its Limits, and How to Tune It

Answering machine detection (AMD) is the system component responsible for classifying each answered call as a live human or a voicemail greeting before connecting the call to an agent. In a predictive dialer workflow, AMD runs in real time between call answer and agent connection—typically with a decision window of 1.5 to 3 seconds. Getting AMD right is critical: false positives (classifying a human as a machine) drop live leads before agents can engage; false negatives (classifying a voicemail as a human) waste agent time and increase abandonment rates.

How AMD Classifies Calls: Signal Analysis

Traditional AMD systems analyze the audio waveform of the first few seconds after call answer, looking for signals that distinguish human greetings from machine greetings: greeting length (voicemail greetings are consistently longer than human "hello"), speech energy patterns, silence duration, and the presence of tones like DTMF or carrier tones that indicate an IVR or answering machine system. Modern AMD systems layer machine learning classifiers trained on large call sample sets to improve accuracy beyond what rule-based signal analysis achieves.

• Greeting duration: The single strongest AMD signal. Human greetings ("Hello?" or "Hello, this is John") are typically under one second. Voicemail greetings ("You've reached the voicemail of John Smith at Acme Corporation, please leave a message after the tone...") run three to ten seconds. AMD systems set a duration threshold and classify calls exceeding it as answering machines.
• Silence detection: The pause after a human greeting is short—humans respond quickly when no one speaks. Voicemail systems have a consistent silence gap between the greeting and the beep. AMD systems use silence duration and placement as a secondary classification signal.
• ML audio classification: Advanced AMD systems use neural network classifiers trained on call audio to identify voicemail greeting patterns, carrier announcement patterns (number not in service, number changed), and fax tones—all of which should result in different dispositions.

Voicemail Drop Integration with AMD

When AMD classifies a call as a machine, the dialer can trigger an automated voicemail drop—playing a pre-recorded message after the beep—rather than simply disconnecting. This turns AMD-filtered machine calls into an additional outreach touchpoint. The compliance requirement: the voicemail drop message must include your organization's name, a callback number, and a disclosure that the message is a marketing communication where applicable under applicable state laws.