Quick answer: Carrier APIs return a verdict you act on: Twilio MachineDetection (Enable or DetectMessageEnd), Telnyx answering_machine_detection (seven values), Plivo, Vonage and Asterisk AMD(). Vapi, Retell, ElevenLabs, Bland and LiveKit decide inside the agent. Published per-call AMD prices on 22 October 2026: Twilio $0.0075, Telnyx $0.002 standard or $0.0065 premium, Plivo $0.0000/min.
There are two layers. Carrier APIs (Twilio, Telnyx, Plivo, Vonage) and the open-source Asterisk PBX listen to the first seconds of an answered outbound call and hand you a verdict string. Agent platforms (Vapi, Retell AI, ElevenLabs Agents, Bland AI) and the LiveKit Agents framework wrap detection inside the voice agent and decide what the agent does next. Every parameter name, value, default and price below was read on that vendor’s own page on 22 October 2026. Why detection misfires is covered in our diagnosis of AI voice agent voicemail detection failures; which threshold pays for your list is worked through in answering machine detection settings by cost.
Where Zian fits
Zian handles answering machines and voicemail on its own calls, so a team running AI appointment setting or AI reception on Zian does not have to pick an AMD mode, tune a millisecond threshold or parse a webhook verdict itself. Zian is an AI sales and reception agents platform, inbound and outbound: it answers incoming calls 24/7 as an AI receptionist, books appointments, and follows up by voice calls, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, in 30+ languages.
This page is a developer comparison, so the vendor tables below rank nothing; they map settings. If you would rather not build the voicemail branch at all, Zian is in a partnership-application beta and each of its three listed plans, including Intel + Widgets (Free Forever), starts with a partnership application. If you want the outbound programme run for you, LeadsNow.ai is the done-for-you option: a managed, pay-per-result AI appointment-setting service that runs on Zian, with leads answered in under 60 seconds, 24/7. One honest limit: Zian publishes no prices during the beta.
How we compared the voicemail detection APIs
Inclusion rule: an API is in if the vendor’s own documentation names a voicemail or answering machine setting you can set on an outbound call. Ten qualified. Evidence rule: every cell comes from the vendor’s own docs, API reference or pricing page, read on 22 October 2026; a price appears only where the vendor’s pricing page prints one. We placed no test calls, so this page publishes no accuracy or speed figure that the vendor did not publish itself. Disclosure: zian.ai publishes this page and Zian appears in the decision table near the end.
Voicemail detection API parameter crosswalk
The table is the whole page in one grid: the switch you set, the timing knobs and their published defaults, the strings you get back, where they arrive, and the published per-call price. “Not printed” means the page we read gives a range but no default.
| Vendor | Switch and values | Timing knobs (published default) | Result values | Where the verdict arrives | Published AMD price, 22 Oct 2026 |
|---|---|---|---|---|---|
| Twilio Programmable Voice | MachineDetection = Enable | DetectMessageEnd; AsyncAmd = true | false (false) |
MachineDetectionTimeout 30 s; MachineDetectionSpeechThreshold 2,400 ms; MachineDetectionSpeechEndThreshold 1,200 ms; MachineDetectionSilenceTimeout 5,000 ms |
human, machine_start, machine_end_beep, machine_end_silence, machine_end_other, fax, unknown |
AnsweredBy and MachineDetectionDuration on the webhook or AsyncAmdStatusCallback |
$0.0075 / call (US) |
| Telnyx Voice API | answering_machine_detection = premium, premium_ios_call_screening_detection, detect, detect_beep, detect_words, greeting_end, disabled (disabled) |
total_analysis_time_millis, greeting_duration_millis and 17 more (eight of them beep_* fields) in answering_machine_detection_config (defaults conflict between pages, see below) |
Standard: human, machine, not_sure. Premium: human_residence, human_business, machine, silence, fax_detected, not_sure |
call.machine.detection.ended, call.machine.greeting.ended, call.machine.premium.* webhooks |
Standard $0.002 per call; Premium $0.0065 per call |
| Plivo Voice API | machine_detection = true | hangup |
machine_detection_time 5,000 ms (2,000–10,000); six advanced fields, ranges only |
Machine = true |
machine_detection_url (POST by default) |
$0.0000/min (US add-on line) |
| Vonage Voice API | Standard machine_detection = continue | hangup; Advanced advanced_machine_detection with mode = default | detect | detect_beep |
beep_timeout (45 in Vonage’s examples) |
status = human | machine; sub_state = beep_start | beep_timeout |
Human / Machine event on the event_url |
Standard: “Free of extra charges”. Advanced: chargeable; Vonage’s guide states “an extra charge of €0.0070 per call”, which we could not confirm on its pricing page (the rate tab did not render for us, 22 October 2026) |
| Asterisk AMD() | Dialplan application AMD(), run after answer |
totalAnalysisTime 5,000 ms; greeting 1,500 ms; initialSilence 2,500 ms; six more |
AMDSTATUS = MACHINE, HUMAN, NOTSURE, HANGUP; plus AMDCAUSE |
Channel variables | Open source, GPL v2; no per-call charge from the project |
| LiveKit Agents | AMD class around the SIP participant |
human_speech_threshold 2.5 s; no_speech_threshold 10 s; timeout 20 s; three more |
human, machine-ivr, machine-vm, machine-unavailable, uncertain |
result.category returned to your agent code |
Not a separate line on the page we read; uses an LLM and STT |
| Vapi | voicemailDetection = "off" or a plan with provider = vapi | google | openai | twilio (off by default) |
backoffPlan.startAtSeconds 5; frequencySeconds 5 (min 2.5); maxRetries 6 (1–10); beepMaxAwaitSeconds 30 |
Call endedReason = voicemail |
End-of-call data; cost itemised as voicemail-detection |
No separate rate on vapi.ai/pricing; itemised per call in the API |
| Retell AI | voicemail_option.action = hangup | static_text | prompt | bridge_transfer; null disables |
Fixed window: first 3 minutes, “isn’t configurable” | disconnection_reason = voicemail_reached or ivr_reached |
Call record and webhooks | No detection fee listed; voicemail calls billed only while the agent is active |
| ElevenLabs Agents | System tool voicemail_detection in built_in_tools |
None published; the LLM decides | Tool call with a required reason string |
Conversation history and batch call results | No separate charge stated on the tool page |
| Bland AI | voicemail.action = hangup (default) | leave_message | leave_message_and_sms | ignore; sensitive (false) |
None published | The action that ran | Call record | Not stated on the pages we read |
Read the last column with care. Twilio and Telnyx both price AMD by the call (“/ call” and “per call”), and neither pricing row says whether a call that is never answered attracts it. Plivo bills an answered call by duration with a 60-second minimum, and its billing table says a call answered by voicemail is billed. That is the reason to hang up fast on machines you do not want.
Twilio AMD: MachineDetection, AsyncAmd and the four timing parameters
Twilio’s switch is MachineDetection, and its two values do different jobs. Enable returns AnsweredBy “as soon as it identifies the called party”, which suits “connect to an agent” logic. DetectMessageEnd returns a human verdict straight away but, for a machine, waits until “the end of the greeting is reached, usually indicated by a beep”, which is what you want before leaving a message. Its published default is “none”, so AMD runs only when you send one of the two values.
AsyncAmd decides whether the callee waits in silence. Twilio’s FAQ says synchronous AMD “introduces several seconds in silence for the callee” and “often leads to hung up calls”; with AsyncAmd=true the call connects at once and the verdict arrives at AsyncAmdStatusCallback. Two limits the FAQ states: asynchronous mode is configurable only on the Calls API, while the Participants API, <Dial><Number> and <Dial><Sip> are asynchronous by default; and Elastic SIP Trunking calls cannot use AMD at all. While it runs, async AMD uses one of the four forked audio streams a call is allowed, shared with Media Streams, SIPREC and Real-Time Transcription.
| Twilio parameter | Allowed values | Default | What raising it does, per Twilio |
|---|---|---|---|
MachineDetectionTimeout |
3–59 s | 30 s | More time to decide; too low returns unknown |
MachineDetectionSpeechThreshold |
1,000–6,000 ms | 2,400 ms | Fewer false machines on long business greetings; slower machine detection |
MachineDetectionSpeechEndThreshold |
500–5,000 ms | 1,200 ms | Catches short voicemails (~2,500 ms suggested); delays human detection by the same amount |
MachineDetectionSilenceTimeout |
2,000–10,000 ms | 5,000 ms | Waits longer on initial silence before returning unknown |
Twilio publishes starting values as well as defaults. Its AMD FAQ and best practices says, for the Enable workflow that connects a human to an agent, that customers calling “individuals, residences, or mobile phones” have had good results with a speech threshold of 1,800–2,000 ms and a speech-end threshold of 1,400–1,500 ms, and for businesses a speech threshold “somewhere between 1800 and 3000”. On speed, it says that on average AMD returns “within ~4 seconds after the call was answered” on default settings. On accuracy it makes the only quantified claim of any vendor here: DetectMessageEnd “is close to 100% accurate in US destinations with default settings”, with accuracy “slightly reduced” internationally, and Enable accuracy “is more relative”. Twilio still says in its own guide that AMD “will not always return the right answer”.
Telnyx answering_machine_detection: seven values and two webhook families
Telnyx sets detection with answering_machine_detection on an outbound dial, or when transferring an inbound call, and its Dial API reference lists seven values with disabled as the default. The four standard modes (detect, detect_beep, detect_words, greeting_end) report on call.machine.detection.ended and, where the mode waits for the greeting, call.machine.greeting.ended. The guide describes detect_words as starting “a 30 second long beep detection” once a machine is detected, and warns that the greeting may still be playing while it counts down.
premium is the mode Telnyx marks “RECOMMENDED”. It reports on call.machine.premium.detection.ended and splits humans into human_residence and human_business. Only two config fields apply to premium, total_analysis_time_millis and greeting_duration_millis. premium_ios_call_screening_detection is the one carrier-layer mode we found, in the documentation read for this comparison, built for Apple Call Screening. After an initial machine verdict it waits for the screening prompt, emits call.machine.premium.greeting.ended with result=prompt_ended so your app can say who is calling and why, or call.machine.premium.call_screening.detected with result=screening, then restarts premium AMD on the screened call. The pricing page lists standard AMD at $0.002 per call and premium at $0.0065 per call. Our Twilio vs Telnyx comparison for AI voice agents covers the per-minute side of that choice.
Plivo and Vonage: the other two carrier APIs
Plivo’s machine_detection takes true (report) or hangup (drop the call), with machine_detection_time defaulting to 5,000 ms in a 2,000–10,000 ms range. Its machine detection guide says detection is asynchronous: on a machine, Plivo posts Machine=true to your machine_detection_url. The Make a Call reference adds six advanced fields, but prints ranges and no defaults: machine_detection_maximum_speech_length 1,000–6,000 ms, machine_detection_initial_silence 2,000–10,000 ms, machine_detection_maximum_words 2–10, machine_detection_initial_greeting 1,000–5,000 ms, machine_detection_silence 500–5,000 ms and machine_detection_answer_time 100–45,000 ms. Plivo’s US pricing page lists Answering Machine Detection at “$0.0000/min”.
Vonage runs two products. Standard machine detection (machine_detection: continue or hangup) is synchronous and, in Vonage’s words, “Free of extra charges”. Advanced machine detection takes three fields: behavior (continue or hangup), mode and beep_timeout. Its default mode is asynchronous, so your first NCCO action plays while detection runs. Vonage says detect behaves like the standard product “with a more accurate identification of machines”, and detect_beep waits for the beep before the next action. Vonage says synchronous identification “takes up to 4-5 seconds” of silence on its side of the call. If you set both products on one call, the advanced one takes precedence.
Asterisk AMD(): the open-source baseline
Asterisk’s AMD() dialplan application, in Asterisk since 1.4.0, is the one engine here you run yourself. The Asterisk docs describe it as detecting answering machines “at the beginning of outbound calls”, called “after the call has been answered (outbound only, of course)”. It reads defaults from amd.conf, which positional arguments override, and sets two channel variables: AMDSTATUS and AMDCAUSE. The cause names the measurement that decided it, such as LONGGREETING, MAXWORDS, INITIALSILENCE or TOOLONG. That makes an Asterisk verdict unusually easy to audit: the docs say the cause carries the measurement and, for most causes, the parameter value it breached.
| AMD() argument | amd.conf default | Crossing it means |
|---|---|---|
initialSilence |
2,500 ms | MACHINE |
greeting |
1,500 ms | MACHINE |
afterGreetingSilence |
800 ms | HUMAN |
totalAnalysisTime |
5,000 ms | Time limit for a decision |
minimumWordLength |
100 ms | Shortest voice counted as a word |
betweenWordsSilence |
50 ms | Silence that separates words |
maximumNumberOfWords |
2 | MACHINE |
silenceThreshold |
256 (scale 0–32,767) | Noise level treated as silence |
maximumWordLength |
5,000 ms | MACHINE |
Telnyx’s standard config uses closely matching snake_case names for eight of these nine numeric arguments, but the defaults differ. Telnyx’s API reference gives 3,500 ms for the greeting and initial silence, five words rather than two, and 3,500 ms for the longest word. Copying an Asterisk tuning across to Telnyx therefore changes behaviour unless you set every field explicitly.
Same knob, different default: the longest-greeting cut
Every audio-timing engine has one parameter that decides how long a greeting can run before it stops sounding human. The engines ship four different defaults for it, against Twilio’s published greeting profile: personal greetings “less than 1800 ms”, business greetings “~1800-3000 ms”, machines “>3000 ms”.
| Engine | Parameter | Default | Past the cut | A 2,500 ms business greeting, on length alone |
|---|---|---|---|---|
| Asterisk | greeting |
1,500 ms | MACHINE | Past the cut |
| Twilio | MachineDetectionSpeechThreshold |
2,400 ms | May be machine | Past the cut, by 100 ms |
| LiveKit Agents | human_speech_threshold |
2.5 s | Deferred to the LLM classifier | At the cut |
| Telnyx standard | greeting_duration_millis |
3,500 ms | Machine | Inside the cut |
| Plivo | machine_detection_initial_greeting |
Not printed (1,000–5,000 ms) | Not described | Depends on your value |
A receptionist’s 2,500 ms business greeting starts as a machine candidate on Asterisk defaults and as human-length on Telnyx defaults, before any other measurement is counted. This is not a claim about end accuracy, since every engine weighs silence and word counts too. It is the reason to set the greeting threshold deliberately for your list. Business lists need a higher cut than mobile lists, and for its Enable agent-connect workflow Twilio publishes separate starting ranges for each.
Rather not tune answering machine detection yourself? Zian’s AI agents handle answering machines and voicemail on their calls, alongside inbound reception and outbound follow-up by voice, SMS, email and WhatsApp. Apply For Partnership
Vapi voicemailDetection: four providers plus a tool
Vapi’s API reference says “By default, voicemail detection is disabled.” You turn it on with voicemailDetection set to a plan whose provider is vapi, google, openai or twilio, or turn it off with the string off. Vapi’s voicemail detection guide recommends its own provider, which it says combines Gemini model-based detection with optional Twilio beep detection. The Vapi, Google and OpenAI plans share a backoffPlan (startAtSeconds default 5, frequencySeconds default 5 with a 2.5 minimum, maxRetries default 6, range 1–10), a type of audio or transcript, and beepMaxAwaitSeconds with a default of 30.
The twilio plan instead passes Twilio’s own fields through, with the same ranges and defaults as the Twilio table above, plus voicemailDetectionTypes, which defaults to machine_end_beep and machine_end_silence. That means a Twilio machine_end_other verdict is not treated as voicemail unless you add it. The message itself is voicemailMessage, text or a .wav or .mp3 URL, and the API says that “If unspecified, it will hang up.” A detected voicemail ends the call with endedReason voicemail, and the cost appears as its own voicemail-detection line, with provider and token counts. The alternative is the Voicemail Tool in model.tools, where the assistant decides. It has a beepDetectionEnabled flag that applies only to Twilio-based calls.
Retell voicemail_option and ivr_option
Retell sets detection on the agent object. Its Create Voice Agent reference takes voicemail_option with an action of hangup, static_text, prompt or bridge_transfer, and an optional detection_prompt of up to 2,000 characters to define what counts as voicemail. Setting the option to null disables it. ivr_option is a separate switch with a hangup action. call_screening_option supplies an agent_identity of up to 100 characters and a call_purpose of up to 300 for when a call screen asks who is calling.
Retell’s voicemail and IVR guide fixes the window: detection runs “during the first 3 minutes after the call connects”, and the window “isn’t configurable”. It runs on outbound phone calls and web calls, the web call being the way to test with a recorded greeting, and not on inbound calls. Outcomes are voicemail_reached or ivr_reached, and Retell says detection “generally adds under 30ms of latency”. One default to watch: new blank voice agents created in the dashboard start with detection on and “Hang up if reaching voicemail” selected, while API-created agents should set voicemail_option explicitly.
ElevenLabs Agents voicemail detection is an LLM system tool
ElevenLabs publishes no timing parameters on its voicemail detection page, because the detector is the agent’s LLM. Its voicemail detection page describes a system tool, voicemail_detection, added under built_in_tools with system_tool_type voicemail_detection: “The LLM analyzes conversation patterns to identify voicemail systems based on automated greetings and prompts”. The tool call carries a required reason string. The only option is an optional voicemail_message, which supports dynamic variables such as {{user_name}}. After detection, and the message if one is set, the call ends, and the event is logged in conversation history and batch call results. To change when it fires, you edit the tool’s description, which is a prompt rather than a threshold.
Bland and LiveKit: the voicemail object and the AMD class
Bland’s voicemail object takes message, an action (hangup by default, leave_message, leave_message_and_sms or ignore), an optional sms block, and sensitive, which “uses LLM-based analysis” and defaults to false. Bland’s call recipes warn that “the action you pass is not always the action that runs”: supplying a message implies leaving one, and ivr_mode: true (or its legacy alias amd) forces ignore. A retry object can redial after a delay with a different voicemail_action.
LiveKit’s answering machine detection is a class you wrap around the SIP participant. It runs once, on the first utterance, pauses agent speech while it listens, and returns one of five categories. machine-unavailable, for a full or unconfigured mailbox, is a category no carrier API above returns. Detection uses an LLM and an STT model, by default google/gemini-3.1-flash-lite and cartesia/ink-whisper through LiveKit Inference, and its timing knobs sit in detection_options. The 20-second timeout is a hard cap only when wait_until_finished is False; with the default True, LiveKit says “a long greeting can run past it”, and because the timer resets when the participant’s audio track is subscribed, “the effective ceiling can reach roughly twice this value”.
Where the vendor docs disagree with themselves
On four vendors’ sites, two of the vendor’s own pages give different answers, or different option lists, for the same field. None of these is a bug report; they are the reason to set every value explicitly rather than trust a default.
| Vendor | Field | Page A says | Page B says |
|---|---|---|---|
| Telnyx | prompt_end_timeout_millis |
AMD guide: default 30,000 ms (1,000–120,000) | Dial API reference: “Defaults to 5000 milliseconds” |
| Telnyx | Premium analysis timeout | AMD guide: “By default, the timeout is set to 30 seconds” | Dial API reference: total_analysis_time_millis default 3,500 |
| Vapi | beepMaxAwaitSeconds |
Guide and field description: 0–60 s | OpenAPI schema: "maximum": 30 |
| Vapi | type |
Guide parameter table: “Only transcript is supported for Google and OpenAI providers” | Same guide’s Detection Types table: audio for all providers |
| Twilio | MachineDetectionTimeout with Enable |
AMD guide: you may decrease it “when MachineDetection=Enabled if you have time constraints” | AMD FAQ: “only relevant for MachineDetection=DetectMessageEnd and shortening this value does not speed up detection for MachineDetection=Enable” |
| Retell | Voicemail actions | Guide: hang up or leave a message | API reference: also bridge_transfer |
Which voicemail detection setting for which job
Pick the job first, then the setting. Each cell is the vendor’s documented route to that job.
| Job | Twilio | Telnyx | Vonage | Plivo | Agent platforms |
|---|---|---|---|---|---|
| Leave a message after the beep | DetectMessageEnd plus AsyncAmd=true; speak on any machine_end_* |
detect_beep or premium; speak on beep_detected |
Advanced, mode detect_beep, behavior continue |
machine_detection=true, then a silence-ended Wait before Speak (Plivo’s sample) |
Vapi voicemailMessage with beepMaxAwaitSeconds; Retell static_text or prompt; ElevenLabs voicemail_message; Bland leave_message |
| Hang up on machines, talk only to humans | Enable; hang up on machine_start |
detect or premium; hang up on machine |
Standard or advanced with hangup |
machine_detection=hangup |
Vapi with no voicemailMessage; Retell hangup; ElevenLabs tool with no message; Bland default |
| A human must not hear silence first | AsyncAmd=true (Calls API only) |
Verdict arrives by webhook after call.answered |
Advanced default mode (asynchronous) |
Asynchronous by design | LiveKit deliberately pauses agent speech while AMD runs |
| Callee may be on Apple Call Screening | No dedicated AMD mode; Twilio’s iOS 26 call screening tutorial pairs AMD with Real-Time Transcriptions | premium_ios_call_screening_detection |
No dedicated mode on the pages we read | No dedicated mode on the pages we read | Retell call_screening_option |
| Warm list, callbacks or inbound | Omit MachineDetection |
Leave at disabled |
Omit both fields | Omit machine_detection |
Vapi off by default; Retell null (dashboard agents start on) |
Leaving a message is a legal question as well as a settings question when the call is telemarketing. Our page on whether ringless voicemail is legal for AI agents covers the consent rules. Those rules turn on how the message is delivered, not on which API detected the mailbox.
Worked cost: 10,000 calls with AMD switched on
Using only the per-call rates printed on each pricing page on 22 October 2026, the add-on for 10,000 calls is: Twilio 10,000 × $0.0075 = $75.00; Telnyx standard 10,000 × $0.002 = $20.00; Telnyx premium 10,000 × $0.0065 = $65.00; Plivo lists $0.0000/min, so $0.00 for the add-on. Carrier minutes come on top in every case, and the other cost is time. A DetectMessageEnd timeout that waits up to 30 seconds for a beep is up to 30 seconds of billed line time on every long greeting. Substitute your own volume; the rates are per call, so the totals scale linearly.
Which voicemail detection API should I use?
| If this is you | Use |
|---|---|
| You want voicemail handled on your calls without building or tuning AMD | Zian, which handles answering machines and voicemail on its calls; for a done-for-you programme, LeadsNow.ai |
| You already place calls on Twilio Programmable Voice | Twilio AMD with AsyncAmd=true, starting from Twilio’s published residential or business values |
| Many callees are on iPhones with Call Screening | Telnyx premium_ios_call_screening_detection |
| You run your own PBX and want no per-call fee | Asterisk AMD(), tuned in amd.conf |
| You build agents on LiveKit | The LiveKit AMD class, which also separates machine-unavailable |
| You use a hosted agent platform | Its native setting: Vapi voicemailDetection, Retell voicemail_option, the ElevenLabs tool, or Bland voicemail |
| The lowest published AMD add-on rate matters most | Plivo, which lists $0.0000/min; Vonage’s standard machine detection is also “Free of extra charges” |
Voicemail detection API questions
What is the best voicemail detection API?
It depends on the layer you build at. If you want voicemail handled without building it, Zian handles answering machines and voicemail on its own calls. If you are building, Twilio publishes recommended starting values and the only quantified accuracy statement here, Telnyx is the only carrier API whose AMD documentation we read describes an Apple Call Screening mode, and agent platforms such as Vapi and Retell expose their own settings. See Telnyx’s AMD guide for the screening mode.
How much does Twilio answering machine detection cost?
Twilio’s US Programmable Voice pricing page lists answering machine detection at $0.0075 per call, as at 22 October 2026, on top of call minutes. Telnyx lists standard AMD at $0.002 per call and premium at $0.0065 per call.
Does Vapi use Twilio for voicemail detection?
Only if you choose it. Vapi’s voicemailDetection accepts four providers: vapi, google, openai and twilio. Vapi says its own provider combines Gemini model-based detection with optional Twilio beep detection. Voicemail detection is disabled by default, according to the Vapi voicemail detection guide and API reference.
Does ElevenLabs have voicemail detection?
Yes. ElevenLabs Agents has a voicemail_detection system tool in which the LLM decides whether a voicemail system answered, with an optional voicemail_message that supports dynamic variables. It exposes no timing thresholds; the ElevenLabs voicemail detection page documents the configuration.
What does AnsweredBy unknown mean in Twilio AMD?
It means Twilio could not decide who answered. The Twilio AMD FAQ says it is returned when the engine lacks enough information, when MachineDetectionTimeout is too short, or when silence outlasts the configured thresholds, and that pushing for faster answers produces more unknown results.
Does voicemail detection work on inbound calls?
Generally no, because nobody’s mailbox answers a call they placed to you. Retell says its detection does not run on inbound calls, and Asterisk describes AMD() as outbound only. Telnyx allows detection when transferring an inbound call, per its AMD guide.
How long does answering machine detection take?
Twilio says that, on average, its AMD returns results within about 4 seconds of answer on default settings, and Vonage says synchronous identification takes up to 4 to 5 seconds. Beep-waiting modes run longer: Twilio’s MachineDetectionTimeout defaults to 30 seconds, per the Twilio AMD guide.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| iOS 26 call screening route: AMD plus Real-Time Transcriptions | Twilio | www.twilio.com/en-us/blog/developers/tutorials/product/detect-ios-call-screening-amd-transcriptions | 22 October 2026 |
| $0.0075 / call, Twilio AMD (US) | Twilio | www.twilio.com/en-us/voice/pricing/us | 22 October 2026 |
| MachineDetection values; AsyncAmd default false; four tuning ranges and defaults (30 s, 2,400 ms, 1,200 ms, 5,000 ms); AnsweredBy values; ~2,500 ms short-voicemail speech-end suggestion; four forked streams | Twilio | www.twilio.com/docs/voice/answering-machine-detection | 22 October 2026 |
| ~4 seconds; close to 100% (DetectMessageEnd, US, defaults); 1,800–2,000 / 1,400–1,500 / 1,800–3,000 ms starting values; greeting profile (<1,800, ~1,800–3,000, >3,000 ms) | Twilio | www.twilio.com/docs/voice/answering-machine-detection-faq-best-practices | 22 October 2026 |
| Standard AMD $0.002 per call; Premium AMD $0.0065 per call | Telnyx | telnyx.com/pricing/call-control | 22 October 2026 |
| Seven answering_machine_detection values; config defaults (3,500 / 800 / 50 / 5 / 256 / 5,000 / 1,500 ms); prompt_end_timeout_millis default 5,000 | Telnyx | developers.telnyx.com/api-reference/call-commands/dial | 22 October 2026 |
| 30-second detect_words beep window; premium timeout 30 seconds; prompt_end_timeout_millis 30,000 (1,000–120,000) | Telnyx | developers.telnyx.com/docs/voice/programmable-voice/answering-machine-detection | 22 October 2026 |
| machine_detection_time 5,000 ms (2,000–10,000); six advanced ranges; 60-second minimum; voicemail calls billed | Plivo | www.plivo.com/docs/voice/api/call/make-a-call | 22 October 2026 |
| Answering Machine Detection $0.0000/min | Plivo | www.plivo.com/voice/pricing/us/ | 22 October 2026 |
| Up to 4–5 seconds synchronous identification; standard detection free of extra charges; advanced €0.0070 per call per the guide; beep_timeout 45 in examples | Vonage | developer.vonage.com/en/voice/voice-api/concepts/advanced-machine-detection | 22 October 2026 |
| AMD() since 1.4.0; AMDSTATUS and AMDCAUSE values | Asterisk project | docs.asterisk.org/Latest_API/API_Documentation/Dialplan_Applications/AMD/ | 22 October 2026 |
| amd.conf defaults (2,500 / 1,500 / 800 / 5,000 / 100 / 50 / 2 / 256 on a 0–32,767 scale / 5,000) | Asterisk project | github.com/asterisk/asterisk/blob/master/configs/samples/amd.conf.sample | 22 October 2026 |
| GPL version 2 licence | Asterisk project | github.com/asterisk/asterisk/blob/master/LICENSE | 22 October 2026 |
| Five categories; 2.5 s / 0.5 s / 1.5 s / 10 s / 20 s / 3.0 s defaults; timeout is a hard cap only with wait_until_finished False; roughly twice timeout | LiveKit | docs.livekit.io/telephony/features/answering-machine-detection/ | 22 October 2026 |
| backoffPlan defaults 5 / 5 (min 2.5) / 6 (1–10); beepMaxAwaitSeconds 30, schema maximum 30; Twilio plan defaults; endedReason voicemail | Vapi | api.vapi.ai/api-json | 22 October 2026 |
| Four providers; beepMaxAwaitSeconds 0–60 | Vapi | docs.vapi.ai/calls/voicemail-detection | 22 October 2026 |
| First 3 minutes; under 30 ms added latency; voicemail_reached / ivr_reached | Retell AI | docs.retellai.com/build/handle-voicemail | 22 October 2026 |
| detection_prompt up to 2,000 characters; agent_identity up to 100; call_purpose up to 300 | Retell AI | docs.retellai.com/api-references/create-agent | 22 October 2026 |
| Voicemail calls billed only while the agent is active | Retell AI | www.retellai.com/pricing | 22 October 2026 |
| voicemail_detection system tool; voicemail_message | ElevenLabs | elevenlabs.io/docs/eleven-agents/customization/tools/system-tools/voicemail-detection | 22 October 2026 |
| voicemail.action default hangup; sensitive default false | Bland AI | docs.bland.ai/api-v1/post/calls | 22 October 2026 |
| 10,000-call add-on totals ($75.00, $20.00, $65.00, $0.00) | Our arithmetic on the Twilio, Telnyx and Plivo rates above | www.twilio.com/en-us/voice/pricing/us | 22 October 2026 |
Ready to hand voicemail handling to the agent? Zian runs AI reception and outbound appointment setting with answering-machine and voicemail handling on calls, in 30+ languages, with HubSpot, Salesforce, HighLevel and Zapier integrations. Apply For Partnership