How Voicemail Detection Works: Signals, AI, and AMD

Written by: Matt Beucler, CEO, Plura AI

Key Takeaways

• Voicemail detection (AMD) analyzes the first few seconds of an answered call using speech duration, pitch, cadence, silence, and beep patterns to classify human vs. machine answers in real time.
• The AMD pipeline follows five stages: answer signal, listen window, feature extraction, classification, and verdict emission, while staying within the FCC’s two-second connection rule.
• Modern AI and LLM systems reduce false positives to 1–3% and classify in 200–800 ms, compared with 15–25% false positives and 1.5–2.5 s latency for legacy rule-based methods.³
• Common failure modes include talkative humans, short voicemail greetings, iOS 26 Call Screening, non-English accents, and IVR or carrier false-answers that can defeat static thresholds.
• Plura AI’s AI Predictive Dialer uses advanced voicemail detection to connect agents only to live humans.

The Voicemail Detection Pipeline: 5 Stages From Answer To Verdict

Every voicemail detection system, regardless of vendor or method, follows the same five-stage pipeline. These stages form the basis for evaluating any AMD implementation.

Stage 1: Answer Signal. The carrier confirms the call was picked up. The SIP layer distinguishes outcomes before audio analysis begins. 180 Ringing means the far end is alerting a real endpoint. 183 Session Progress means early media is flowing, where carrier announcements, SIT tones, and most false-answers arrive. 200 OK is the answer from which analysis starts.

Stage 2: Listen Window. The system collects audio for analysis. Documented start delays range from 1.5 to 3 seconds across implementations. The listen window is where detection speed is won or lost. A system that needs 4 seconds of audio cannot match one that classifies in 840 milliseconds.

Stage 3: Feature Or Transcript Extraction. The system extracts acoustic features such as speech duration, silence gaps, pitch, and cadence, or it transcribes the audio to text. Audio-based detection works in any language without speech recognition. Transcript-based detection looks for semantic cues like “please leave a message.”

Stage 4: Classification. A rule-based engine, a machine learning model, or an LLM classifies the audio as human, voicemail, IVR, or unknown. Rule-based systems compare measured audio events against hard thresholds. AI systems output a confidence score instead of a binary rule.

Stage 5: Verdict Emission. The system routes the call based on the classification by connecting an agent, playing a voicemail drop, or hanging up. The verdict must arrive within the compliance budget. Under U.S. telemarketing regulations (47 CFR 64.1200(a)(7)), a call is treated as abandoned if it is not connected to a live sales representative within two seconds of the called person’s completed greeting.²

See how AI-powered voicemail detection connects your agents only to live humans in a live demo.

How Voicemail Detection Uses 4 Core Audio Signals

Voicemail detection systems analyze four primary audio signals to distinguish a human answer from a machine. Each signal carries documented thresholds that vary across implementations.

Audio Feature Extraction: Speech Duration, Pitch, And Cadence

Systems analyze the initial audio for duration, pitch, tone, cadence, and background noise. Human greetings typically last 500–2,000 milliseconds, while voicemail greetings commonly extend 3,000–15,000 milliseconds or longer. Machine learning models like Wave2Vec, fine-tuned for AMD, analyze the first 2–4 seconds of audio and achieve accuracy rates of 97–98.5% by recognizing subtle patterns in audio spectrograms.

Greeting Length Analysis: The Primary Classification Signal

A human usually says a short “Hello?” under 2 seconds and then stops to listen. A voicemail plays a long, uninterrupted greeting or automated message such as “You’ve reached…”. Genesys Cloud treats under 2,200 ms of speech followed by 700 ms or more of silence as a live person. Twilio exposes similar parameters with a 2,400 ms speech threshold and a 1,200 ms speech-end threshold.⁴ Greeting length alone remains fragile. Genesys’s documentation notes that “Hello, this is John Smith” reads as human, while “Hello, this is John Smith. Thank you for calling Widgets Inc, how can I assist you today?” reads as voicemail, so a talkative receptionist can defeat silence-pattern rules.

Silence Thresholds: The Pause That Tells The Story

Humans speak briefly and then wait in silence for a reply. Machines have different patterns of continuous speech and rolling pauses. After a human says “Hello,” they typically pause 700–2,000 milliseconds waiting for a response. Voicemail systems continue playing recorded messages with minimal silence until the end-of-greeting beep. Asterisk’s built-in AMD function, used by VICIdial, configures initial_silence at 2,500 ms (max silence before classifying as MACHINE), greeting at 1,500 ms (max greeting length before MACHINE), and after_greeting_silence at 800 ms (silence after greeting to confirm MACHINE).

Tone And Beep Recognition: The Definitive Machine Signal

Listening for specific electronic sounds like a recording beep confirms a voicemail box is ready for a message. Typical beeps are 800–1,200 Hz lasting 200–500 milliseconds. Detection uses frequency analysis to identify these tones. Twilio splits machine outcomes into machine_end_beep, machine_end_silence, and machine_end_other because many systems never beep at all and instead use silence or verbal prompts like “Please leave a message.” Beep detection is nearly 100% accurate but slow, since the beep only comes after the full greeting plays, often 5–15 seconds. Modern systems therefore combine beep detection with faster classification methods.

Rule-Based Vs. AI/LLM Voicemail Detection: The 2026 Shift

The industry is shifting from rule-based heuristics to AI and LLM-powered classification. This change affects accuracy, latency, and how teams manage outbound performance.

Rule-Based/Heuristic Systems: The Legacy Approach

Older systems like Asterisk’s app_amd measure exact milliseconds of speech and silence against hardcoded thresholds. The decision logic works in four steps:

1. If there is silence for initial_silence ms after the call is answered, it is likely a machine.
2. If more than maximum_number_of_words words are spoken without a long pause, it is a voicemail greeting.
3. If continuous speech exceeds greeting ms, it is a machine playing a recorded message.
4. After speech stops, if silence exceeds after_greeting_silence ms, it confirms a machine.

These static rules misclassify 15–25% of calls as false positives. They also degrade over time. Fixed thresholds cannot adapt to changes in voicemail greetings, iOS call screening, carrier-level STIR/SHAKEN filtering, and modern voicemail systems. Rule-based systems that scored well in 2015 lose accuracy without code changes.

AI And LLM Models: The Modern Approach

Modern tools process audio streams in rolling windows or transcribe the first words instantly. They use semantic cues, such as “please leave a message,” to classify calls accurately within two seconds. AI AMD models analyze audio across a rolling temporal context window of typically 200–800 milliseconds and output a confidence score instead of a binary rule. If confidence exceeds 90%, the system makes a decision. If confidence remains low, the analysis window extends by another 200 ms for edge cases.

The measurable improvements are significant. AI and LLM systems cut false positives from 15–25% down to 1–3% and classify calls in under a second, compared with 1.5–2.5 seconds for many rule-based systems.

Metric	Rule-Based (e.g., Asterisk AMD)	AI/LLM-Based
False positive rate	15–25%	1–3%
Classification latency	1,500–2,500 ms	200–800 ms
Adaptability to accents/languages	Static thresholds	Retrained on regional data
Handling of iOS 26 Call Screening	Misclassifies as voicemail	Can distinguish screener prompts

A 20-agent call center making 500 connected calls per hour at a 20% AMD false positive rate drops 100 live humans per hour. At a 3% rate, only 15 are dropped, recovering 85 live conversations per hour, or 680 per 8-hour shift.³

Why Voicemail Detection Fails: 5 Documented Failure Modes

No voicemail detection method reaches 100% accuracy. Distinguishing a human “Hello?” from a machine greeting in a fraction of a second is inherently probabilistic. These five failure modes appear most often in production.

Failure Mode 1: Talkative Humans. A human answering with a long greeting like “Hello, this is Ahmed speaking, how can I help you?” can be misclassified as voicemail by word count and length. Genesys’s documentation acknowledges this failure case with receptionists.

Failure Mode 2: Short Voicemail Greetings. A short voicemail greeting like “Leave a message” can be mistaken for a live human. Custom business greetings that start with a human voice are acoustically indistinguishable from a live answer in the first 400 ms.

Failure Mode 3: iOS 26 Call Screening. Apple’s Call Screening automatically answers calls from unknown numbers, asks for the caller’s name and reason, and transcribes the reply before the phone rings. AMD systems misclassify the screener as a human because it plays a long voice segment with multiple words, then goes silent while waiting. With iOS holding 59% US market share and over 70% adoption of iOS 26 within 6 months of release, this creates a significant practical problem.

Failure Mode 4: Non-English Greetings And Regional Accents. Regional greeting patterns, non-English voicemail systems, and elderly callers with slower cadence fall outside the training distribution of many classifiers. Different SIP carriers process audio differently through noise reduction, echo cancellation, and level normalization, and accents vary across markets.

Failure Mode 5: IVR Systems And Carrier False-Answers. IVR menus that list options such as “Press 1 for sales, press 2 for support” can be misclassified as voicemail. Carrier false-answers (FAS), where the network itself answers the call, account for about 14% of answered calls across one AMD provider’s network.

The Speed Vs. Accuracy Tradeoff In AMD

Voicemail detection always balances speed against accuracy. Faster detection reduces dead air for human callers but increases misclassification risk. Slower detection improves certainty but increases the chance that humans hang up during silence.

Verdict speed varies by method:

• Silence-pattern rules: 3–30 seconds
• Audio classifiers: under 4 seconds
• Transcript-plus-LLM methods: around 1 second

LiveKit reports a median time to decision of 840 milliseconds for transcript-based classification. Bland fine-tuned Wave2Vec on the first two seconds of call audio and reported 98.5% accuracy.⁴ Twilio’s detection timeout defaults to 30 seconds.

The tradeoff has direct compliance implications. AMD latency directly affects abandonment-rate exposure because a call only counts as “connected” once routed past AMD to an agent. A longer AMD detection window adds to the answer-to-connect delay measured against the FCC’s 3% abandonment cap.² A dialer that over-paces without accounting for AMD’s classification window can breach abandonment limits.

Configuration recommendations vary by use case:

• High-volume campaigns: startAtSeconds: 1–2, frequencySeconds: 2.5, maxRetries: 4–5, detectionTimeout: 10–15 seconds
• High-value leads: startAtSeconds: 2.5–3, frequencySeconds: 3–4, maxRetries: 7–10, detectionTimeout: 20–30 seconds

Watch the AI Predictive Dialer handle the speed-accuracy tradeoff in a live outbound campaign.

Voicemail Detection In Practice: Plura AI’s AI Predictive Dialer

Voicemail detection is the engine that makes modern predictive dialers work. Plura AI’s AI Predictive Dialer uses voicemail detection to eliminate voicemails and connect agents only to live humans. The dialer decides who to call next using stateful conversion signals such as historical answer rates, prior negotiation outcomes, and prior offer-acceptance bands, then routes every answered call through voicemail detection before an agent speaks.

The economics are direct. In high-volume outbound operations where 40–60% of answered calls are voicemails, the difference between an AMD system at 75% accuracy and one at 97% accuracy translates into significant agent-hours either wasted or recovered.³ Plura’s AI Predictive Dialer runs on Plura’s own FCC-licensed audio bridging carrier, which avoids third-party CPaaS dependencies and enables branded caller ID, STIR/SHAKEN authentication, and real-time DNC scrubbing at the carrier level before every call leaves the network.¹

As noted earlier, the difference between a 20% and 3% misclassification rate recovers 85 live conversations per hour for a 20-agent call center. Run your numbers through Plura’s ROI calculator to check your cost savings in real time.

Experience AI-powered voicemail detection on a live outbound campaign with a Plura demo.

Frequently Asked Questions

How Many Rings Before Voicemail Is Normal?

By default, a phone rings for around 20 seconds before voicemail takes over, which equals roughly four to six rings because one ring is about five seconds. Voicemail forwarding timers can typically be set in 5-second increments (5, 10, 15, 20, 25, or 30 seconds), with 30 seconds as the maximum. Most callers tolerate between three and five rings, approximately 15 to 20 seconds, before hanging up. After that point, abandon rates climb steeply. For outbound campaigns, this ring window becomes the practical budget AMD has to classify the call before the voicemail system picks up and the detection problem changes entirely.

Why Does Voicemail Detection Fail?

Voicemail detection fails for five documented reasons:

• Talkative humans whose long greetings look like machine messages
• Short voicemail greetings that sound like live answers
• iOS 26 Call Screening, which plays a recorded prompt and waits, often misclassified by rule-based systems
• Non-English greetings and regional accents outside the classifier’s training distribution
• IVR systems that list options like a recorded message

Rule-based systems are particularly vulnerable because they compare audio against fixed thresholds that cannot adapt to new greeting patterns. AI-based systems reduce these failure modes but do not eliminate them, since AMD classification remains probabilistic.

What Is The Difference Between Rule-Based And AI Voicemail Detection?

Rule-based systems such as Asterisk’s app_amd measure exact milliseconds of speech and silence against hardcoded thresholds like initial_silence, greeting, after_greeting_silence, and maximum_number_of_words. They misclassify 15–25% of calls and degrade over time as greeting patterns change. AI and LLM systems analyze audio across a rolling temporal context window of 200–800 milliseconds, output a confidence score instead of a binary rule, and can be retrained on new data. These systems reduce false positive rates to 1–3% and classify faster, typically 200–800 ms vs. 1,500–2,500 ms for many rule-based approaches.

How Does Voicemail Detection Work On Twilio?

Twilio’s outbound AMD is controlled by the machineDetection mode on calls.create().⁴ The webhook result includes AnsweredBy values such as human, machine_start, machine_end_beep, machine_end_silence, fax, and unknown. These values determine whether you connect an agent, leave a message, hang up, or retry. Key parameters include MachineDetectionTimeout (default 30 seconds), MachineDetectionSpeechThreshold (2,400 ms), and MachineDetectionSpeechEndThreshold (1,200 ms). Twilio’s documentation acknowledges the algorithm does not always classify correctly across all humans and greetings. Twilio also offers a Voicemail Detection operator in Conversation Intelligence Classic for post-call classification, which is separate from real-time AMD gating.

How Accurate Is Voicemail Detection In 2026?

Vendor-published accuracy ranges from 94.7% overall (LiveKit) to 98.5% (Bland), but these numbers are measured on each vendor’s own audio. For example, LiveKit reports 95.7% F1 on human, 98.2% on IVR, and 97.3% on voicemail. In contrast, rule-based systems typically achieve 60–85% accuracy in real-world conditions, while AI and ML-based systems reach 93–99%. Vendor accuracy numbers may overstate real-world performance on production traffic with different carriers, greetings, and accents. When evaluating any AMD vendor, ask for the false positive rate, the detection window used, and whether accuracy holds under production load. A well-trained model on under-provisioned infrastructure can perform poorly when CPU utilization exceeds 70–80%.

---

Read Next

• What Is Voicemail Detection in Voice AI?
• Voicemail Beep Detection: How It Works and Why It Fails
• How to Improve Voicemail Detection Accuracy: 2026 Guide
• Predictive Dialer Voicemail Detection: How AMD Works
• Voicemail Detection Best Practices for AI Predictive Dialers