Answering machine detection, or AMD, is the technology that analyzes the first seconds of an answered outbound call to decide what picked up: a person, a voicemail system, a fax machine or an automated attendant. A dialer uses it to connect only live humans to the rep, and to release the line or drop a message in every other case.
You will also see it called human voice detection or voicemail detection. This article explains how it works, where it gets things wrong, what it costs in delay, and why it becomes essential as soon as you dial several numbers at once.
What exactly is answering machine detection?
When an outbound call is answered, the phone network only signals that something picked up. It doesn’t say what. A person, a cell phone voicemail, a desk phone’s mailbox, an IVR or a fax machine all trigger the same answer signal.
Answering machine detection fills that gap. It listens to the start of the call and returns a verdict, which almost every provider labels the same way:
| Verdict | What answered | What the dialer does |
|---|---|---|
| Human | A person | Connects the rep |
| Machine | Voicemail, answering machine | Releases the line or drops a message |
| Fax | A fax machine | Releases the line, flags the number |
| Unknown | Ambiguous audio, timeout | Depends on settings, often treated as human |
Verdicts documented by Twilio (human, machine_start, fax, unknown), Asterisk (HUMAN, MACHINE, NOTSURE) and Telnyx (human, machine, not_sure), checked on November 5, 2026.
Telnyx recommends treating the “unknown” verdict as if a human answered: passing a voicemail to a rep costs a few seconds, hanging up on a prospect costs a conversation.
How does classic, cadence-based detection work?
The historical method doesn’t understand what is being said. It measures when someone speaks and for how long. It rests on a simple observation: a person who answers says a word or two, then stops to let the caller talk. A voicemail plays a greeting several seconds long without waiting for anyone.
The AMD module in Asterisk, the open source PBX that many dialers are built on, makes these rules easy to read. Here are its default settings, as published in its configuration file:
| Asterisk setting | Default value | Rule |
|---|---|---|
| Maximum initial silence | 2,500 ms | If nobody speaks for 2.5 s, verdict machine |
| Maximum greeting length | 1,500 ms | A longer greeting, verdict machine |
| Silence after greeting | 800 ms | If the greeting is followed by 0.8 s of silence, verdict human |
| Maximum number of words | 2 | More than 2 words in the greeting, verdict machine |
| Total analysis time | 5,000 ms | No conclusion after 5 s, verdict unknown |
Source: amd.conf.sample file from the Asterisk project, checked on November 5, 2026.
Twilio, one of the main programmable voice providers, describes the same principle: its AMD “isolates human speech audio and measures periods between speech and silence in the greeting”. Its defaults are a little more tolerant: speech lasting more than 2.4 seconds tips the verdict toward machine, and 1.2 seconds of silence after speech leads to a human verdict.
Its weakness: it judges behavior, not a voice. Any human who talks like a machine, and vice versa, fools it.
What do AI models add to human voice detection?
Newer systems no longer just time the silences. They analyze the audio content: timbre, recording quality, intonation, background noise, the presence of a beep, and sometimes the words themselves through speech recognition. A model trained on real calls recognizes a greeting even when it is short, and a person even when they introduce themselves at length.
A team of researchers published an approach on arXiv in 2024 based on a recurrent neural network that processes the audio stream in real time. They report over 96% accuracy on their test set, and over 98% with an added silence detection step. Telnyx, for its part, offers a “premium” mode that combines speech recognition and machine learning, and even distinguishes a human answering at home from a human answering at a business.
False human, false machine: what errors does AMD make?
Talking about “false positives” and “false negatives” quickly gets confusing, because it all depends on what you call positive. Twilio uses two clearer terms, which we use here.
| Error | What happens | Consequence |
|---|---|---|
| False human | A voicemail is classified as a person | The rep hears a recorded greeting, hangs up, loses a few seconds |
| False machine | A person is classified as a voicemail | The prospect answers, the line is released or a message plays to them: the conversation is lost |
These two errors don’t cost the same: a false machine wastes a prospect who actually picked up, the scarce thing the dialer is supposed to produce. On a B2B list where only a small minority of dials reach a person (see our cold call connect rate benchmarks), every false machine hurts.
The problem is that the two errors move in opposite directions, as Twilio’s documentation says plainly:
- Raising the allowed speech duration reduces false machines on long business greetings, but slows down voicemail detection; lowering it does the reverse.
- Raising the silence required after speech helps with short voicemails, but delays human detection: going from 1,200 to 2,500 ms adds about 1.3 seconds of delay, in Twilio’s example.
So there is no perfect setting, only a trade-off. For B2B outbound, the sensible trade-off is almost always to protect the humans: better for the rep to hit a voicemail now and then than to lose a prospect who answered.
How much time does detection add before the connection?
This is the hidden cost of AMD. To conclude that a person answered, the system has to hear them speak, then stop. That silence is, by definition, time the prospect spends waiting.
It is the gap everyone has experienced: you say “Hello?”, and nothing for a second or two. The longer it lasts, the more obvious the automated dialer, and the harder the opening.
To shorten it: an engine that decides quickly on short answers, a fast connection once the verdict is in, and a rep ready to talk the instant the call arrives. We cover these settings in our parallel dialing best practices.
How does AMD relate to voicemail drop?
Automatically dropping a pre-recorded voicemail is the natural next step after detection. Once a voicemail has been identified, the dialer can play a message recorded in advance while the rep moves on to the next call.
But dropping a message requires a second detection: the end of the greeting. If the message starts while the greeting is still playing, its opening is lost and the prospect hears a sentence cut in half. Providers offer dedicated modes for this: Twilio reports the end of a greeting detected “by beep”, “by silence” or “other”; Telnyx offers beep detection and end-of-greeting detection based on a long silence.
A false machine is more visible here than anywhere else: the prospect who answered hears a recorded message that doesn’t let them speak. One more reason to tune detection on the cautious side.
For the message itself, see our guides on leaving a cold call voicemail, B2B voicemail scripts and how to record voicemail templates.
Why is AMD essential in a parallel dialer?
A parallel dialer dials several numbers at once for a single rep and connects the first one who answers. As soon as that first answer is connected, the other lines in the round are cancelled while they are still ringing, and those prospects see a missed call.
Without detection, that mechanism falls apart: the first thing to “answer” would often be a voicemail, passed to the rep while the other lines, perhaps one with a real person, were cancelled.
With detection, every answered line is classified in real time. Too many false humans, and the rep keeps getting interrupted by voicemails while the other lines were cancelled for nothing. Too many false machines, and prospects who answered are lost without anyone noticing. Too much delay, and they have already hung up.
So line count isn’t everything: opening more lines multiplies the answers to classify and the cancelled calls. We show in how many lines a parallel dialer should run why 4 lines are enough for most B2B lists. To compare tools that combine multi-line dialing and detection, see our roundup of the best parallel dialers.
Business voicemail or cell phone: why does accuracy vary?
Not all voicemails sound alike, and not all people answer the same way.
| Situation | What trips detection | Likely error |
|---|---|---|
| Person answering at a business: “Acme Logistics, this is Sarah speaking, how can I help?” | Long greeting, several words, no pause | False machine |
| Short cell phone voicemail: “It’s Paul, leave a message” | Brief greeting, followed by silence before the beep | False human |
| Company switchboard or IVR: “Press 1 for…” | Neither a human nor a personal voicemail | Classification varies by tool |
| Person who answers and stays silent, or in a noisy place | Long initial silence, or noise mistaken for speech | False machine, or unknown |
| Smartphone screening of unknown callers | A voice assistant answers on the person’s behalf | Machine, although the person might have taken the call |
Business greetings are the costliest case in B2B: someone who answers with the company name and their own often speaks for more than two seconds without a pause, the textbook profile of a voicemail to an engine that times speech. With switchboards and IVRs there is nobody to reach directly, and the right behavior depends on what you want to do with those numbers.
The last case is recent. Since iOS 26, the iPhone can screen unknown numbers: the caller is asked for their name and reason for calling before the phone even rings, according to Apple. To a dialer, that is an automated voice answering, and Telnyx has added a dedicated detection mode. A detection engine has to be maintained, not just installed.
What are the limits of answering machine detection?
AMD is an aid, not a guarantee. Three limits are worth keeping in mind, on top of the delay covered above.
No accuracy is 100%. Even the best published models get a few calls in a hundred wrong: dozens of misclassified prospects across thousands of calls a month.
Accuracy is measured on your calls. Cell phones or landlines, language, country, share of switchboards and time of day all change the results. Regularly listen to a sample of recordings classified as “machine”: it is the only way to spot false machines, which are otherwise invisible.
Detection doesn’t replace a clean list. A file full of switchboards, dead lines or off-target cell phones will generate lots of “machine” verdicts, whatever the engine.
How does Skipcall handle voicemail detection?
Skipcall’s power dialer and parallel dialer include AI voicemail detection. It recognizes voicemail greetings with a stated 95% accuracy, along with invalid numbers and switchboards. When it isn’t a person, the line is released without the rep hearing the greeting, and dialing carries on. The first live answer is passed to the rep, and the other lines in the round are cancelled.
The same detection triggers voicemail drop: a pre-recorded message, or an SMS depending on the sequence settings, is left automatically once the voicemail is reached, while the rep moves on to the next call. The call is placed normally and the contact’s phone rings: this is not ringless voicemail. The dropped message is logged on the contact record, then in the CRM (HubSpot, Pipedrive, Salesforce or Attio).
Skipcall’s parallel dialer runs up to 4 lines and is included in the Starter plan, from $70 per month per license billed annually (see pricing).
Skipcall isn’t the right choice for everyone. If you are looking for a voice agent that talks in place of your reps, an IVR or inbound call routing, Skipcall does none of that: its AI determines who answered and triggers the drop, it doesn’t hold any conversation. And if your team makes few calls to a short list of highly targeted accounts, a multi-line dialer with detection may not be worth it: our which dialer quiz helps you decide.
What to remember
- Answering machine detection (AMD) decides, on every answered call, whether it is a person, a voicemail, a fax or an unknown case, using silence cadence or AI models.
- False human and false machine move in opposite directions. In B2B outbound, protect the humans: a false machine costs a conversation.
- Detection imposes a short delay, around one to two seconds with published settings, and voicemail drop depends on it entirely.
- In a parallel dialer, detection quality matters more than line count.
- Measure accuracy on your own calls, business greetings and cell phone voicemails included.
Sources checked on November 5, 2026: Twilio Answering Machine Detection documentation; Asterisk AMD application documentation and its amd.conf.sample file; Telnyx Answering Machine Detection documentation; K. Altwlkany et al., “A Recurrent Neural Network Approach to the Answering Machine Detection Problem”, arXiv, 2024; Apple Support, “Manage unknown callers on iPhone”.
