Best Answering Machine Detection for VICIdial in 2026: Built-In AMD, Khomp, AMDY and VM Hunter Compared

If you run VICIdial, you have four realistic choices for answering machine detection in 2026: the built-in Asterisk detector, the Khomp gateway integration, and two hosted AI services, AMDY and VM Hunter. They work in different ways, sit in different places in the call path, and fail on different calls.

This is a comparison written by one of the four, so read it accordingly. We describe the other three from their own public documentation and posts, link to it, and do not publish accuracy figures for products we have not measured. Where another option is the better fit, we say so. At the end is a test you can run in a day to get your own answer.

What a detector has to handle in 2026

Ten years ago the job was "person or voicemail". Today an answered call can be any of these:

1. A live person, usually saying one or two words and waiting.
2. A voicemail greeting, personal or the carrier's default.
3. Call screening. iOS 26, Google Call Screen and call-blocking handsets answer with a prompt while the person reads a transcript.
4. A disconnected-number message, answered by the carrier instead of a SIP error.
5. An IVR or business menu.
6. Nothing: dead air, or no audio because of a trunk fault.
7. Tones: fax, a beep, ringback after a false answer.

A detector that only separates the first from the rest loses value in two places. Screening calls are reachable people. And disconnected numbers and no-audio calls need different handling from voicemail: retire the first, fix the trunk for the second.

The questions to ask of any option are therefore: which of these does it label, how fast, can I check its verdicts, and what does it change in my call path?

Option 1: the built-in detector (Asterisk AMD)

How it works. VICIdial sends answered calls to a dialplan extension (8369 on a stock system) that runs Asterisk's AMD() application. It measures the timing of sound and silence against eight thresholds: initial silence, greeting length, the silence after a greeting, number of words and so on. It returns HUMAN, MACHINE or NOTSURE with the timing rule that fired.

Strengths.

• Free, already installed, and nothing leaves your server.
• Fully under your control. Every threshold is yours to set.
• Good at the clear cases: long greetings and long silences.

Limits.

• It does not know what was said. "Hello?" then a pause and "Hi, this is Mike" then a pause are the same to it.
• Call screening, carrier messages and voicemail are all just MACHINE.
• No-audio calls are INITIALSILENCE, the same as a quiet person.
• Operators who measure it carefully typically find 15 to 20% of live answers classified as machines on consumer lists. The figure moves a lot with the list and the tuning.

Best for. Small operations, business lists where people answer with a clear greeting, and anyone who has not yet tuned it. Start there: our guide to VICIdial AMD settings covers each value.

Option 2: Khomp

How it works. Khomp is a call classification product with an official VICIdial integration. VICIdial sends the SIP INVITE to a Khomp gateway, which forwards it to your carrier through its own routing. When the carrier answers, the gateway analyses the audio before passing the answer back, and returns a conclusion and a pattern that VICIdial acts on. Details are on the VICIdial integration page.

Strengths.

• Classification happens before the call reaches the dialer, in the SIP path, so it sees signalling as well as audio. That is a good position for catching false answers and carrier messages.
• Supported directly by the VICIdial Group.
• An established product with years in production.

Things to weigh.

• Your calls route through the gateway, which is a change to your call path and carrier routing.
• Requirements and pricing are by quotation; contact VICIdial or Khomp.
• Ask how call screening is labelled and what evidence you get per call. We could not confirm either from public material.

Best for. Larger VICIdial clusters that want detection in the network path and a vendor relationship through the VICIdial Group.

Option 3: AMDY

How it works. AMDY is a hosted AI detector that, according to its site, classifies the acoustic signature of the audio and does not transcribe speech. A client on the dialer sends the audio and receives a status.

What its site states.

• Detection starts in 1/8 of a second.
• Because it does not depend on words, it is language-agnostic.
• Tones, fax, dead air and false answers are classified acoustically.
• It flags honeypot or spam-trap numbers.
• 23 analytics reports and a per-detection log.
• A free Sandbox plan of 50,000 detections a month, and a one-command install for VICIdial, Asterisk and FreeSWITCH.

In posts on the VICIdial forum, AMDY has said that iOS call screening and Google Voice are tagged with a status such as MACHINE:GVOICEAMD.

Things to weigh.

• "Starts in 1/8 of a second" is when analysis begins. Ask what the time from answer to verdict is on live calls, and measure it on yours.
• Screening calls arrive as a machine status. Check whether your dialplan can route them to an agent.
• An acoustic detector has no transcript, so ask what evidence you get when you dispute a verdict.

Best for. Multilingual lists, where a detector that does not depend on vocabulary has a real advantage, and teams that want honeypot flagging and a large free tier.

Option 4: VM Hunter

How it works. VM Hunter is a hosted detector that reads the words as well as the sound. An EAGI script on the dialer streams the callee's audio from the moment of answer. The service transcribes the first two seconds, analyses the audio for beeps, tones, static and silence, and returns AMDSTATUS and AMDCAUSE, the same two variables the built-in detector sets. If the callee is still talking at two seconds and the words are not yet decisive, it listens for up to three.

Strengths.

• It tells you why. Causes include voicemail wording, beep, call screening, disconnected number, silence and no audio.
• Call screening is its own label, CALLGUARD_PHRASE, so the dialplan can play a short intro and connect an agent.
• Evidence for every call: transcript, reason and an audio sample in the portal, with CSV export.
• No-audio calls are separated from quiet lines in the call log, which exposes trunk faults.
• Nothing is added to the SIP path. Install is one script and one dialplan extension, about 30 minutes.
• Free plan of 5,000 calls a month; paid plans from $100 a month. The clients are open source.

Limits, stated plainly.

• It is tuned for English. A Spanish voicemail is still classified as a machine from the audio, but a Spanish-speaking person is not recognised by their words.
• Audio is streamed to a hosted service. If policy requires everything on-premise, that rules it out.
• A verdict takes about two seconds. That is deliberate, and we explain why below, but a product that decides sooner will hand off sooner.
• A person who answers and says nothing is reported as silence, as with every detector.

Best for. English-language consumer and business lists, operations that want to audit verdicts, and anyone losing leads to call screening.

Side by side

From public documentation as of October 2026. "Check" means we could not confirm it from public material.

Sound or words: what our own test showed

The central difference between the two hosted AI options is method. One side of that argument says words are slow and unnecessary; the other says sound alone cannot separate a person from a greeting that starts the same way.

We tested the question. We built an acoustic model ourselves, from a dataset of more than 500,000 recordings, and ran it against VM Hunter on 4,000 calls that agents had dispositioned. Every call was the hard case: a lone "Hello" that an older build of our engine had passed as human.

On the 2,271 calls where speech began in the first second:

• The acoustic model passed 78.7% of the answering machines to an agent. VM Hunter passed an estimated 62.3%.
• Both passed about 94 to 95% of live people.
• At a one-second decision point, the acoustic model lost 21% of live people.
• 61% of those machines were caught by neither.

Three cautions. That acoustic model is ours, not AMDY's product, which may do better or worse. The VM Hunter figure is an offline estimate. And this is the hardest slice of traffic, not the general mix. The full write-up has the method, and the acoustic model is open source if you want to run it yourself.

Which should you choose?

• You have not tuned the built-in detector: do that first. It is free and may be enough.
• You want detection in the network path with VICIdial Group support: talk to them about Khomp.
• Your lists are multilingual: an acoustic detector such as AMDY has the advantage. Test it.
• Your lists are English and you are losing people to "Hello?" drops or call screening: test VM Hunter.
• You cannot decide: you do not have to. Both hosted services have free tiers, and each uses its own routing extension, so they can run side by side.

The one-day test

1. Split one list at random into campaigns, one per detector, including your current one.
2. Same agents, same hours, recordings on.
3. Next morning, listen to the first three seconds of 200 calls from each campaign. Mark what was really there.
4. For each detector count people dropped, machines passed to agents, and screening calls and what happened to them.
5. Compare time from answer to agent, and talk time per agent hour.

Vendor numbers, ours included, come from the vendor's own traffic. Yours will differ, and one day of your own calls is worth more than any comparison page, including this one.

To set up the VM Hunter side, follow the VICIdial guide. The individual comparisons are on the comparison page: Khomp, AMDY, and the carrier-based options from Twilio, Telnyx and Plivo.