How to Configure Voicemail Detection Timeout Settings

Written by: Matt Beucler, CEO, Plura AI

Updated 2026-09-10

Key Takeaways

• Voicemail detection timeout settings differ across platforms. Twilio and Vapi default to 30 seconds, while LiveKit uses 20 seconds and several vendors do not publish their values.⁴
• The four threshold categories (overall detection timeout, speech threshold, speech end or silence threshold, and initial silence timeout) govern different phases of call audio and require separate configuration.
• Short timeouts create three failure modes: more unknown results, broken message-end detection, and live humans routed to voicemail drops.
• Unit discipline matters. Twilio uses seconds for the overall timeout and milliseconds for speech and silence thresholds, so unit mistakes cause severe misconfiguration.
• Plura AI’s AI Predictive Dialer handles voicemail detection natively on its own FCC-licensed carrier, which removes much of the parameter-guessing required by API-wrapper tools. See how Plura handles voicemail detection in a live demo.

How to Read the Voicemail Detection Timeout Settings Comparison Table

The table below maps each platform to its primary answering machine detection (AMD) timeout parameter, its published default, and its valid range. Parameter names appear as inline code. Every value traces to that vendor’s published documentation. When a platform’s default is not published in public documentation, the table calls that out directly.

Platform	Parameter Name	Default	Valid Range
Twilio	MachineDetectionTimeout	30 seconds	3–59 seconds
Vapi	machineDetectionTimeout	30 seconds (mirrors Twilio AMD defaults)	Not independently published; delegates to Twilio AMD
Retell	Not published in public documentation	Not published	Not published
LiveKit	Not published as a single named parameter	20-second timeout; 2.5-second human speech threshold	Not published as a discrete range
Pipecat	Pipeline-level; no single vendor-defined parameter	Not published	Not published; configured at the transport or pipeline layer
Amazon Connect	RingTimeoutInSeconds (ring timeout); AMD via AnswerMachineDetectionConfig	60 seconds (ring timeout default); AMD timeout not published	15–60 seconds (ring timeout)
Bland AI	Not published in public documentation	Not published	Not published

Compare plans and rates side by side.

How to Configure the Four Voicemail Detection Threshold Categories

Twilio’s AMD documentation defines four distinct threshold categories. Each category controls a different phase of the call audio, so mixing them up creates avoidable errors.

Overall Detection Timeout

This setting is the hard deadline for the AMD decision. On Twilio, MachineDetectionTimeout defaults to 30 seconds, with a valid range of 3–59 seconds. If the deadline fires before a verdict, Twilio returns AnsweredBy=unknown. LiveKit’s equivalent timeout defaults to 20 seconds. Vapi mirrors Twilio’s 30-second default when routing through Twilio AMD. Retell, Pipecat, Bland AI, and Amazon Connect’s AMD-specific timeout do not publish equivalent values.

Speech Threshold (~2400 ms)

This parameter defines the minimum duration of continuous speech that triggers a machine classification. On Twilio, MachineDetectionSpeechThreshold defaults to 2400 ms, with a valid range of 1000–6000 ms. Speech shorter than the threshold may classify as human, while longer speech may classify as machine. Vapi’s published configuration examples use the same 2400 ms default. LiveKit’s equivalent is 2.5 seconds. Retell, Pipecat, Bland AI, and Amazon Connect do not publish this value.

Speech End or Silence Threshold (1200–5000 ms)

This parameter defines how much silence must follow a speech segment before the system treats that segment as complete. On Twilio, MachineDetectionSpeechEndThreshold defaults to 1200 ms, with a valid range of 500–5000 ms. Twilio recommends increasing this to about 2500 ms to handle short voicemail greetings that pause mid-greeting, and notes that higher values add equivalent delay to human detection. Vapi’s published examples use 1200 ms. Other platforms do not publish this value.

Initial Silence Timeout (3000–5000 ms)

This parameter defines how long the system waits for any speech before returning an unknown result due to dead air. On Twilio, MachineDetectionSilenceTimeout defaults to 5000 ms, with a valid range of 2000–10000 ms. Vapi’s published examples use 5000 ms. Twilio recommends 3000 ms when the goal is to return unknown quickly after silent pickups. Other platforms do not publish this value.

How to Diagnose a Voicemail Detection Timeout That Is Too Short

Once the four threshold categories are clear, the next step is understanding what breaks when the overall detection timeout is set too low. A MachineDetectionTimeout set too low produces three observable failure modes.

First, premature unknown classifications increase. Twilio’s documentation confirms that a low timeout value produces more unknown AMD results because the engine stops listening before it has enough audio for a confident decision. unknown functions as a timeout signal rather than a third classification category.

Second, when using MachineDetection=DetectMessageEnd, a short timeout breaks message-end detection. Twilio states that MachineDetectionTimeout is only relevant for DetectMessageEnd, and that a low value may return unknown even when a machine answered because the timeout fires before the greeting and beep complete.

Third, live calls can be misrouted. A machine_start call lasting only a few seconds is the signature of a misrouted human. A person answers, the detection classifies them as a machine, a voicemail drop starts, and the person hangs up while the call still bills as completed.

How to Set Twilio MachineDetectionTimeout

On Twilio, the parameter is MachineDetectionTimeout, expressed in seconds. The default and valid range are covered above. Values above 60 are reduced to 60, and invalid non-numeric values default to 60. The parameter is set on the outbound call create request alongside machineDetection (which accepts Enable or DetectMessageEnd), MachineDetectionSpeechThreshold, MachineDetectionSpeechEndThreshold, and MachineDetectionSilenceTimeout.

Twilio notes that 30 seconds is sufficient for most residential voicemail boxes but frequently not enough for business voicemail boxes, and that most configuration changes observed in production are decreases rather than increases.

How to Configure Vapi and Retell Voicemail Detection Timeout

Vapi’s voicemail detection routes through Twilio AMD. One published Vapi configuration example uses machineDetectionTimeout: 30, mirroring Twilio’s AMD default. The same example sets machineDetectionSpeechThreshold: 2400, machineDetectionSpeechEndThreshold: 1200, and machineDetectionSilenceTimeout: 5000. Vapi’s own documented default is not confirmed. Vapi’s current recommended path for voicemail detection is LLM-based detection via function calling, with the Twilio AMD path marked legacy. Parameter names match Twilio’s convention and use the same units.

Retell does not publish AMD timeout parameter names, defaults, or ranges in its public documentation. According to third-party analysis, Retell treats voicemail and IVR detection as a first-class agent setting that runs within a timeout window. Engineers configuring Retell should consult Retell’s support documentation or account team for current defaults.

How to Configure LiveKit and Pipecat Voicemail Detection Settings

LiveKit does not ship a single named voicemail detection timeout parameter equivalent to Twilio’s MachineDetectionTimeout. LiveKit’s AMD implementation returns one of five values: human, machine-ivr, machine-vm, machine-unavailable, or uncertain. It uses a 2.5-second human speech threshold and a 20-second timeout by default. LiveKit runs a fast heuristic path for short greetings followed by silence and falls back to a language-model classifier when a transcript requires reasoning. Configuration details beyond these defaults should be confirmed from LiveKit’s telephony documentation at docs.livekit.io/agents/start/telephony/.

Pipecat is an open-source framework. Pipecat does not expose a single vendor-defined AMD timeout parameter. Voicemail detection behavior is implemented at the pipeline and transport layer. On Pipecat, a VoicemailDetector runs a parallel pipeline with a classifier whose only job is to output CONVERSATION or VOICEMAIL, gating TTS output until the decision lands. Timeout behavior is configured as part of the pipeline logic rather than through a discrete parameter. Engineers should consult Pipecat’s documentation at docs.pipecat.ai for current pipeline configuration options.

How to Tell Speech Threshold and Silence Timeout Apart

Across these platforms, two parameters cause the most confusion because they sound similar but govern opposite phases of the call. These two parameters are frequently conflated.

Speech threshold (MachineDetectionSpeechThreshold on Twilio) measures the duration of continuous speech after the call connects. If speech runs longer than the threshold, the system classifies the answering party as a machine. If speech is shorter, it may classify as human. This parameter fires during active speech. As noted earlier, the default is 2400 ms on Twilio.

Initial silence timeout (MachineDetectionSilenceTimeout on Twilio, default 5000 ms) measures how long the system waits for any speech to begin after the call connects. If no speech appears within the silence timeout window, the system returns unknown. This parameter fires during silence before any speech starts.

The interaction works as follows. If a call connects and the answering party says nothing for 5 seconds, the silence timeout fires and returns unknown before the speech threshold is evaluated. If the answering party speaks immediately, the silence timeout is bypassed and the speech threshold takes over. Twilio advises using MachineDetectionSilenceTimeout set to 3000 ms when the goal is to return unknown quickly after silent pickups, because the two parameters address different failure modes.

How to Troubleshoot Voicemail Detection Timeout Settings

The matrix below connects common AMD symptoms to likely root causes and the specific parameter to adjust. Use it as a quick diagnostic guide when your AMD distribution shows unexpected patterns.

Observed Symptom	Root Cause	Parameter to Change
High share of unknown results	Detection timing out before verdict	Raise MachineDetectionTimeout; raise MachineDetectionSpeechThreshold
Agent connects to dead air after voicemail pickup	Timeout too short; greeting not finished before deadline	Raise MachineDetectionTimeout to 45+ seconds for residential
Voicemail message starts mid-greeting, cutting off first words	Beep not detected before timeout; DetectMessageEnd deadline firing early	Raise MachineDetectionTimeout; verify MachineDetection=DetectMessageEnd is set
Live humans routed to voicemail drop (short machine_start calls)	Mode problem: Enable deciding too early on borderline greetings	Switch to MachineDetection=DetectMessageEnd or enable AsyncAmd=true
Calls returning unknown after silent pickup	Initial silence exceeds silence timeout	Lower MachineDetectionSilenceTimeout to 3000 ms to return unknown faster
Short voicemail greetings classified as human	Speech threshold too high relative to greeting length	Lower MachineDetectionSpeechThreshold to 1500–2000 ms; raise MachineDetectionSpeechEndThreshold to 2000–2500 ms
Business receptionists classified as machines	Long human greeting exceeds speech threshold	Raise MachineDetectionSpeechThreshold to 3000 ms for business-line campaigns

Compare native detection to parameter-based AMD in a live Plura demo.

How to Keep Unit Discipline for Seconds and Milliseconds

The most common misconfiguration in voicemail detection is entering a value in the wrong unit. On Twilio, the four AMD parameters use two different units in the same API call. The list below shows each parameter, its unit, and its default and range.

• MachineDetectionTimeout: seconds (default 30, range 3–59)
• MachineDetectionSpeechThreshold: milliseconds (default 2400, range 1000–6000)
• MachineDetectionSpeechEndThreshold: milliseconds (default 1200, range 500–5000)
• MachineDetectionSilenceTimeout: milliseconds (default 5000, range 2000–10000)

The two units create two opposite failure modes. Entering 30000 for MachineDetectionTimeout (treating it as milliseconds) will be capped at 60 seconds by Twilio, not interpreted as 30 seconds. Entering 30 for MachineDetectionSpeechThreshold (treating it as seconds) sets a 30 ms speech threshold, which will classify nearly every call as human regardless of greeting length.

Amazon Connect’s RingTimeoutInSeconds is expressed in seconds, consistent with Twilio’s overall timeout convention. LiveKit’s published defaults are expressed in seconds for the overall timeout and seconds for the speech threshold. Always verify the unit in the vendor’s parameter reference before setting a value.

How to Balance Speed, Accuracy, and Cost

A detector tuned to be confident is a detector that is late. That tradeoff is a policy decision rather than a tuning decision. The timeout value you set determines where you land on three competing axes: detection speed, classification accuracy, and billing cost per call. The table below shows recommended timeout bands for common use cases and the rationale behind each.

Use Case	Recommended Timeout Band	Rationale
Fast-connect AI agents (outbound lead response)	Roughly 2–4 seconds, with detection windows kept under about 3 seconds	Prioritizes speed, accepts higher unknown rate, protects live-answer experience
Residential voicemail outreach	Default 30 seconds is probably sufficient for most cases	Covers varied personal greeting lengths and reduces unknown and mid-greeting drops
Business-line outreach	10–15 seconds	Business greetings tend to be brief, so shorter timeouts reduce idle billing seconds
High-value lead campaigns (accuracy priority)	20–30 seconds	Accepts slower detection for 97–98.5% accuracy when false positives cost more than delay3
International calling	25–35 seconds	Handles wider variation in voicemail implementations and carrier behavior across countries

Twilio bills $0.0075 per call for AMD on every dial it is enabled for, whether or not a machine answers³. A longer timeout does not change per-call AMD cost, but it extends the window during which a connected call accrues per-minute charges. Keeping voicemail drop clips short, around fifteen seconds per the cited source, helps keep most voicemail calls inside a single billed minute, since carriers round each call up to the next whole minute.

How to Handle DetectMessageEnd and Beep Detection

MachineDetection=DetectMessageEnd is the mode used for leaving a voicemail after the beep rather than mid-greeting. It waits for the machine greeting to finish before returning a verdict. The possible return values are machine_end_beep, machine_end_silence, machine_end_other, human, fax, and unknown.

A short MachineDetectionTimeout breaks this mode directly. Twilio states that when MachineDetection=DetectMessageEnd, low values for MachineDetectionTimeout may return unknown even if a machine answered and the engine could have determined so, because the timeout triggers before the end of the machine greeting. In that case, Twilio may return machine_end_other as the fallback.

Twilio notes that 30 seconds is frequently not enough for business voicemail boxes when using DetectMessageEnd. For campaigns targeting business lines with long greetings, Twilio documents increasing MachineDetectionTimeout within its allowed range as the remedy.

How to Adjust for iOS 26 Call Screening

iOS 26 introduced expanded call screening that intercepts calls from unfamiliar numbers before they ring through to the recipient. A LiveKit Agents user running thousands of answered calls per day reports that around 30% of their calls are voicemail or iOS 26 Call Screening bots³, and that share has been climbing as more recipients turn screening on.

iOS 26 screening creates a new AMD edge case. The call is answered by a screening bot rather than a human or a carrier voicemail system. The screening bot may ask the caller to state their name or reason for calling before deciding whether to pass the call through. This interactive automated response does not match the audio patterns of either a standard human greeting or a carrier voicemail greeting, so standard AMD thresholds tuned for those two cases may misclassify screened calls.

The practical effect on timeout settings is straightforward. Screened calls may produce longer initial audio than a human greeting but shorter than a full voicemail greeting, landing in the ambiguous zone around the speech threshold. Increasing MachineDetectionSpeechThreshold toward 3000 ms and using DetectMessageEnd rather than Enable reduces the risk of classifying a screening interaction as a live human pickup. Platforms that communicate with iOS 26’s call-screening layer at the carrier level can present a recognizable caller identity to the screening system, which converts screened calls into pickups before AMD runs.

How Plura AI Handles Voicemail Detection Timeout Settings

All of the platform-specific configuration above assumes you are managing AMD through third-party APIs. Plura AI takes a different approach. Plura’s AI Predictive Dialer runs voicemail detection natively on Plura’s own FCC-licensed audio bridging carrier, not through a third-party CPaaS. This architectural difference affects timeout configuration. Because Plura owns the carrier layer, detection, branded caller ID issuance, and real-time DNC (Do Not Call) scrubbing are enforced inside the platform before dial. This differs from API-wrapper tools, where these functions are bolted on through parameters passed to a third-party network.

For operations teams currently managing MachineDetectionTimeout, MachineDetectionSpeechThreshold, and related parameters across Twilio-based API resellers, Plura’s native detection layer reduces the parameter-guessing those tools require. The platform handles the detection decision internally, and timeout behavior is managed at the carrier level rather than through per-call API configuration.

Plura’s AI Predictive Dialer is a Vici Dial alternative and supports multi-line dialing. It shares a Stateful Conversation Database with Plura’s AI SMS, AI RCS, and AI Webchat channels, so context from a prior text or web interaction is available to the dialer on the next outbound call. STIR/SHAKEN (Secure Telephone Identity Revisited / Signature-based Handling of Asserted information using toKENs) caller ID authentication runs on every outbound call.¹ Plura supports TCPA (Telephone Consumer Protection Act) and DNC compliance workflows at the platform level on every outbound contact.^1,2 Customers remain responsible for their own compliance obligations and should consult qualified counsel regarding their specific regulatory requirements.

Watch a live demo of carrier-level detection with Plura.

Frequently Asked Questions (FAQ)

What Happens If the Voicemail Detection Timeout Is Too Short?

A timeout set too low produces the three failure modes described earlier: more unknown results, broken message-end detection, and misrouted live calls. The observable symptom of the third failure is a cluster of very short machine_start calls in your AMD distribution.

Does DetectMessageEnd Change the Timeout?

DetectMessageEnd does not change the numeric value of MachineDetectionTimeout, but it changes what that timeout governs. In Enable mode, the timeout bounds how long the engine waits to make any human, machine, or unknown decision. In DetectMessageEnd mode, the timeout bounds how long the engine waits for the greeting to finish before it can return a machine_end_beep, machine_end_silence, or machine_end_other verdict. Because DetectMessageEnd must wait for the entire greeting to complete, the same 30-second default that is sufficient in Enable mode is frequently insufficient for business voicemail boxes with long greetings. Twilio’s documentation notes that almost all configuration changes observed in production are decreases rather than increases.

What Is the Difference Between Speech Threshold and Silence Timeout?

Speech threshold (MachineDetectionSpeechThreshold on Twilio) measures the duration of continuous speech after the call connects. If speech runs longer than the threshold, the system classifies the answering party as a machine. If speech is shorter, it may classify as human. This parameter fires during active speech.

---

Read Next

• Voicemail Detection Parameters: Cross-Platform Guide
• How Vapi Handles Voicemail Detection: A Configuration Guide
• Voicemail Drop After Detection: The 2026 Setup Guide
• How to Improve Voicemail Detection Accuracy: 2026 Guide
• Voicemail Detection Integration for Contact Centers