Key takeaways
- Twilio charges $0.0075 per call for answering machine detection on US calls (Twilio voice pricing, checked November 2026).
- Twilio’s detection times out after 30 seconds by default and then reports “unknown”; the timeout can be set from 3 to 59 seconds.
- Twilio’s documentation says AMD “will not always return the right answer”, so every setup needs a plan for misclassified calls.
- With synchronous AMD the caller hears silence while detection runs, which on a sales call can count toward the two-second abandoned-call rule.
What is answering machine detection (AMD)?
Answering machine detection is an automatic check, run as an outbound call connects, that labels the answer as a person or a machine, and it works on timing and sound, not on understanding words. When the call connects, the detector measures how long the first stretch of speech lasts and how long the silence after it lasts. People tend to answer with a short greeting and wait; recorded greetings tend to run on for several seconds. Detectors also listen for fax tones and, if asked, for the beep or the silence that ends a greeting.
Every detector makes two kinds of mistake. A false positive labels a person as a machine, so the dialer hangs up on someone who answered. A false negative labels a machine as a person, so a rep is connected to a greeting. Tightening the settings to catch more machines raises the first error, and loosening them raises the second, which is why platforms expose thresholds instead of a single on-off switch.
How does answering machine detection work on Twilio?
On Twilio, you turn AMD on per call with the MachineDetection parameter, through the Calls API or a <Number> inside <Dial> (Twilio AMD docs). “Enable” returns a result as soon as Twilio decides, with the values human, machine_start, fax or unknown. “DetectMessageEnd” waits until a machine’s greeting ends and reports machine_end_beep, machine_end_silence or machine_end_other, so a message can start after the beep.
By default detection is synchronous: Twilio holds the call until it decides. With AsyncAmd set to true, the call continues and the result arrives on a callback. Four settings tune it: the timeout (30 seconds by default, 3 to 59 allowed), the speech threshold (2,400 ms), the speech-end threshold (1,200 ms) and the silence timeout (5,000 ms). Twilio bills $0.0075 per call that uses it.
| AnsweredBy value | Mode | Meaning | Typical dialer action |
|---|---|---|---|
| human | Both | A person is likely on the line | Keep the rep connected |
| machine_start | Enable | A greeting started | Offer skip or a live voicemail |
| machine_end_beep | DetectMessageEnd | The greeting ended with a beep | Start a message now |
| machine_end_silence | DetectMessageEnd | The greeting ended in silence, with no beep heard | Start a message, or let the rep listen first |
| machine_end_other | DetectMessageEnd | The greeting ended some other way Twilio could not classify | Let the rep listen before speaking |
| fax | Both | Fax tones detected | Hang up and mark the number bad |
| unknown | Both | No decision before the timeout | Treat as a person |
How accurate is answering machine detection?
No reliable public accuracy figure exists, because results depend on the list, the carriers, local greeting habits and the settings, and vendors do not publish comparable tests. Twilio itself says its AMD “will not always return the right answer” and reports how long each detection took in a MachineDetectionDuration field, which you can use to measure your own results.
The practical test is your own data. Record a sample of calls, compare the AMD label with what you hear, and count both errors. If people are being hung up on, raise the speech threshold or stop auto-hanging up; if reps keep landing on greetings, lower it. A setting that shows the rep the result and lets them decide limits the damage of either error.
How does AMD affect the abandoned-call rule?
AMD interacts with US telemarketing rules on silence. The FCC treats a call as abandoned when it is not connected to a live sales representative within two seconds of the person’s completed greeting, and caps abandoned calls at 3% of those answered by a person per campaign (47 CFR 64.1200(a)(7)). Synchronous detection that keeps a person waiting, or a dialer that hangs up on people it mislabels as machines, can push a campaign toward that limit. Predictive dialers rely on AMD most, which is part of why predictive dialing needs careful tuning to keep abandoned calls in check. This is general information, not legal advice.
What mistakes do teams make with answering machine detection?
Most AMD problems come from automating the decision without checking the errors it makes, because every hang-up on a person and every rep stuck on a greeting costs something. Watch for these.
- Auto-hanging up on every “machine” result without sampling recordings, so real people are dropped.
- Leaving synchronous detection on for live calls, so people hear several seconds of silence and hang up.
- Treating “unknown” as a machine: a person who says nothing for a moment gets dropped.
- Using AMD to play recorded sales messages without checking consent rules and your carrier’s terms on voicemail drops.
- Ignoring the cost: at $0.0075 a call, 10,000 dials a month add $75 to the Twilio bill.
When is answering machine detection not worth it?
AMD helps most when a rep would otherwise sit through many greetings, such as cold lists with a low connect rate. It adds less when a team calls warm leads who often pick up, when reps want to leave a personal voicemail on every no-answer anyway as part of a multi-touch follow-up plan, or when each call is valuable enough that a wrongly dropped person costs more than a few seconds of greeting. For those teams, a preview step and a voicemail script the rep reads live do the same job without the per-call charge.
How does answering machine detection (AMD) relate to Rapidialer?
Rapidialer’s Detect voicemail setting is off by default; owners and admins switch it on for the whole organization in dialing settings. It runs Twilio AMD asynchronously on the <Number>, so the rep hears the person at once and the result appears a few seconds later as “Voicemail detected” with Leave voicemail and Skip. A second option, Skip voicemails automatically, hangs up and logs the outcome as Voicemail. In the opt-in parallel mode (2 to 5 lines per rep), machines are detected and hung up on the same way, so only people reach the rep. Rapidialer never plays recorded voicemail or sales messages (besides your optional recording-disclosure notice, the only other audio is a short company-name line for a person answered after the rep was taken in parallel mode), because Twilio says voicemail drops do not fit its terms (Twilio Support). Detection is billed per call on your own Twilio account, and the power dialer moves on after the rep saves an outcome. See Twilio’s per-minute rates for the rest of the bill.
Related terms
- Predictive dialerA dialer that calls several numbers per available agent, predicting when agents will be free.
- Abandoned callAn answered outbound call with no agent on the line within two seconds of the greeting.
- TwiMLTwilio Markup Language: XML instructions that tell Twilio what to do on a call or message.
Answering machine detection (AMD): questions and answers
How much does Twilio charge for answering machine detection?
Twilio’s US voice pricing page lists answering machine detection at $0.0075 per call, checked November 2026. The charge applies to each call where detection is requested, on top of the per-minute rate for the call itself. At 2,000 detected calls a month that is $15, so it is small next to minutes, but worth including in a cost estimate.
What is the difference between Enable and DetectMessageEnd?
Both are Twilio MachineDetection modes. Enable reports as soon as Twilio decides whether a person or a machine answered, which suits dialers that connect a rep or skip quickly. DetectMessageEnd reports people at once but waits until a machine’s greeting finishes, reporting a beep, silence or other ending, so a message can start at the right moment.
Why do people hear silence when a dialer calls them?
Usually because the dialer is waiting before it connects a rep: a predictive dialer looking for a free agent, or synchronous answering machine detection holding the call while it decides. Asynchronous detection avoids the second cause, since the rep hears the person immediately and the result arrives a few seconds later without holding the audio.
Can AMD detect voicemail on mobile phones?
Yes, it listens to the audio, so it works the same way on mobile and landline numbers. Results vary with how greetings sound: carrier default greetings, short personal greetings and visual voicemail setups can each change the timing AMD relies on. Sample recordings from your own lists to see how often it gets mobile calls right.
Is answering machine detection legal?
Detecting voicemail is legal, but what you do around it is regulated. In the US, a sales call where nobody is connected within two seconds of the greeting counts as abandoned, and abandoned calls are capped at 3% per campaign. Playing a recorded sales message after the beep is a prerecorded call that needs prior express written consent. Ask a lawyer about your setup.
Does AMD make reps faster?
It can, on lists where many calls reach voicemail, because reps skip greetings instead of listening to them. The gain shrinks when answer rates are high or when reps leave a voicemail on most calls anyway. Measure conversations per hour before and after switching it on, and check recordings for people wrongly skipped, before deciding it helps.