Voicemail Detection Accuracy: 2026 Benchmarks and Tuning

Written by: Matt Beucler, CEO, Plura AI

Updated September 2026

Key Takeaways

• Modern AI-powered voicemail detection reaches 95-99% accuracy, while legacy rule-based systems achieve 60-85% in real-world conditions.³
• False positives, where live humans are misclassified as voicemail, are the costlier error because dropped conversations cannot be recovered. Target rates under 5%.
• Configuration tuning such as adjusting speech thresholds, start delays, and using ML-based classifiers can improve accuracy by 15-25% over default settings.
• Carrier-level branded caller ID and FCC-licensed infrastructure reduce spam labels and improve answer rates before voicemail detection even runs.
• Teams seeking carrier-grade AMD at scale can book a live demo with Plura to see how it integrates into a compliant outbound stack.

Voicemail Detection Benchmarks for 2026

Published accuracy figures span a wide range depending on the detection method, model architecture, and the audio dataset used for evaluation. The table below synthesizes figures from named sources. Every data point is measured on the vendor’s own audio unless otherwise noted, so treat these as directional, not apples-to-apples comparisons, as Cekura’s 2026 analysis explicitly cautions.

System / Model	Published Accuracy	Detection Window	Source
Bland AI fine-tuned Wave2Vec4	98.5%	First 2 seconds	Bland AI (Feb 2025)
Bland AI CNN (Mel spectrograms)	97%	First 4 seconds	Bland AI (Feb 2025)
LiveKit transcript-based	97.3% F1 (voicemail)	840 ms median	LiveKit via Cekura (2026)
Twilio AMD (ML-based)4	~94%	~4 seconds avg	Twilio via NextLevel.AI (2026)
arXiv temporal-feature system	96.1% combined	46 ms inference	arXiv:2604.09675 (2026)
Legacy rule-based AMD	60–85%	4–5 seconds	Belsmart, Murf AI (2026)

The arXiv:2604.09675 system is notable because it includes both benchmark and production validation. Across 77,000 live calls, it maintained a 0.3% false positive rate and a 1.3% false negative rate. These figures reflect operational conditions rather than controlled test sets.

The CallSphere cascade approach, which layers Twilio AMD with an LLM second pass, reports ~96% production accuracy versus ~83% for AMD-only configurations. This illustrates how multi-signal architectures close the gap between lab and production performance.

Operational Causes of Voicemail Detection Errors

AMD errors fall into two categories with different operational costs.

False positives occur when a live human is misclassified as a machine. The dialer disconnects or plays a voicemail drop to a real person. The industry target is a false positive rate under 5%, while many legacy systems run 10-20%. The arXiv production system achieved 0.3% in 77,000 calls.³ False positives are the costlier error because a dropped live conversation is gone permanently.

False negatives occur when a voicemail greeting is misclassified as a human. An agent or AI voice agent waits through the greeting. The industry target is under 10%, with each occurrence wasting 5-10 seconds of agent time. Many operations accept a higher false negative rate specifically to protect against false positives.

Several factors push real-world accuracy below controlled-test figures:

• Custom greetings: Conversational voicemail openings like “Hey, thanks for calling!” challenge systems trained on standard carrier messages.
• Audio quality: Network latency, background noise, and codec compression obscure the acoustic signals used for classification.
• Carrier and device diversity: Mobile carriers introduce connection delays, compress audio differently, and run varied voicemail systems. These differences explain why real-world accuracy often trails controlled-test figures.
• IVR systems and call screening: Apple’s iOS 26 Call Screening automatically answers calls from unknown numbers. It asks for a name and reason, then transcribes the reply before the phone rings. This behavior can inflate connect rates and cause AMD systems to misclassify the screening bot as a human.
• Detection speed: Faster detection uses less audio for the decision, which increases error rates.

Error rates also depend on how many calls reach voicemail in the first place. 67% of Americans say their general practice is to not answer calls from unknown numbers but to check a voicemail if one is left. One consumer survey found that 90% of consumers are uncomfortable answering unidentified calls and that 78% have missed an important call in the last month because they did not answer an unidentified number.

AMD accuracy and caller ID reputation are therefore linked. Calls that present with a company name rather than an unknown number reach more live humans in the first place, which reduces the share of calls that go to voicemail. TransUnion states that customers are up to 105% more likely to answer a branded call.

How to Improve Voicemail Detection Accuracy

Configuration tuning delivers measurable gains on any platform. Organizations that systematically test with representative call samples and iteratively adjust parameters typically achieve 15-25% accuracy improvements over default settings, regardless of provider.

Key levers:

• Timeout thresholds: Twilio’s detection timeout accepts 3-59 seconds with a default of 30 seconds. Longer timeouts improve decision quality but increase dead air before routing.
• Speech threshold tuning: Twilio uses a 2,400 ms speech threshold and a 1,200 ms speech-end threshold. Genesys Cloud treats under 2,200 ms of speech followed by 700 ms of silence as a live person. Increasing the speech threshold to 3,000-4,000 ms reduces false positives on verbose human greetings.
• Start delay (startAtSeconds): Increasing startAtSeconds to 3-4 seconds reduces premature classifications caused by initial silence or connection noise.
• Greeting length analysis: Human greetings run 500-2,000 ms, while voicemail greetings run 3,000-15,000+ ms. This duration difference is the primary classification signal.
• Beep detection: 800-1,200 Hz tones lasting 200-500 ms signal voicemail end-of-greeting. Because not all voicemail systems use beeps, rely on beep detection only as a secondary signal, not the primary classifier.
• Model selection: ML-based classifiers such as Wave2Vec or CNNs over Mel spectrograms outperform rule-based heuristics on custom greetings and edge cases. For example, modern transcription-based AMD reaches 95-98% accuracy with sub-three-second latency, compared to 70-85% for legacy energy-and-silence heuristics.
• Cascade approach: Combining carrier AMD with an LLM second pass achieves ~96% production accuracy versus ~83% for AMD-only configurations. The LLM pass only triggers on ambiguous cases (~12% of calls), which keeps cost overhead negligible.

The Speed vs. Accuracy Trade-Off

Every AMD system faces the same constraint: faster detection means less audio to analyze, which increases error rates. The right balance depends on campaign type.

For high-volume campaigns, aggressive configurations such as startAtSeconds of 1-2 and detectionTimeout of 10-15 seconds maximize throughput. For high-value leads, conservative configurations such as startAtSeconds of 2.5-3 and detectionTimeout of 20-30 seconds maximize accuracy, even if detection takes 5-8 seconds rather than 2-3 seconds.

The latency cost is real. Each additional second of post-answer silence drops conversation rates by 5-8%. Regulatory context also sets a hard deadline.² Under 47 CFR 64.1200(a)(7), one FCC rule treats a telemarketing call as abandoned if it is not connected to a live sales representative within two seconds of the called person’s completed greeting.² This rule sets the detect-and-route budget for many outbound campaigns. Teams should consult qualified counsel on how AMD configuration interacts with their specific regulatory obligations.

An AMD accuracy number must be paired with its detection window and false positive rate to be meaningful. A vendor that quotes one figure without the other two presents an incomplete view of performance.

Frequently Asked Questions

What Is a Good Voicemail Detection Accuracy Rate?

Modern AI-powered voicemail detection reaches 95–99% accuracy in published benchmarks, while legacy rule-based systems achieve 60–85% in real-world conditions. As noted earlier, the headline accuracy only matters alongside the false positive and false negative rates. A system at 97% accuracy with a 10% false positive rate can perform worse operationally than one at 95% accuracy with a 2% false positive rate because false positives represent dropped live conversations.

Many teams use a false positive rate under 5% and a false negative rate under 10% as baseline operational thresholds. For high-value campaigns, leaders often prioritize false positive rate over headline accuracy and use conservative configurations with longer detection windows to protect live connections.

Why Does Voicemail Detection Keep Misclassifying Human Answers?

The most common causes are a speech threshold set too low, detection starting too early, or rule-based heuristics that cannot handle custom greetings. Humans with longer greetings get flagged as machines when thresholds are too aggressive. Initial connection silence can be misread as a machine pause when detection starts too early.

Practical fixes include increasing startAtSeconds to 3–4 seconds to allow humans more time to begin speaking naturally and increasing speechThreshold to 3,000–4,000 ms to accommodate verbose greetings such as business receptionists. Switching from rule-based to ML-based detection also improves performance on custom greetings. As covered in the tuning section, some teams add a cascade approach with an LLM second pass on ambiguous cases to recover calls that single-signal AMD misroutes.

Is Faster Voicemail Detection Always Better?

Faster detection reduces the audio available for analysis, which increases error rates in both directions. High-volume campaigns may accept 2–3 second detection windows to maximize throughput and accept a higher error rate as a tradeoff. High-value operations often prioritize accuracy with 5–8 second detection windows, even at the cost of slightly longer call handling times.

Regulatory context also sets a practical ceiling on detection speed. Some FCC rules treat a telemarketing call as abandoned if it is not connected to a live representative within two seconds of the called person’s completed greeting. AMD must complete within that window on many compliant campaigns. Teams should consult qualified counsel on how their specific AMD configuration interacts with applicable regulations.

How Does iOS 26 Call Screening Affect Voicemail Detection?

Apple’s iOS 26 Call Screening automatically answers calls from unknown numbers, asks the caller for their name and reason for calling, and transcribes the reply before the phone rings through to the recipient. From an AMD perspective, this behavior creates a new classification challenge. The system hears a voice responding to a prompt, which can be misclassified as a human answer when it is actually a screening bot.

This pattern inflates apparent connect rates while corrupting reachability data. Platforms that rely on acoustic heuristics alone are most vulnerable. Transcript-based or LLM-based classifiers that can identify screening language patterns handle this edge case more reliably. Branded caller ID that presents a company name to the screening layer also reduces the share of calls that trigger screening in the first place.

What Is the Difference Between Synchronous and Asynchronous AMD?

Synchronous AMD holds the call in silence until the detection verdict arrives before connecting to an agent or AI voice agent. This approach maximizes decision certainty but creates 2-5 seconds of dead air that human callers often interpret as a robocall and hang up.

Asynchronous AMD connects the call immediately and runs detection in the background. This eliminates the silence penalty but requires the platform to handle a mid-conversation branch if voicemail is detected after the agent has already started speaking. For AI voice agents, asynchronous detection is the standard production approach because it maintains conversational flow for live humans while still catching voicemail in the background. The tradeoff is that late detections require a graceful mid-call transition rather than a clean pre-connect branch.

Why Plura AI for High-Volume Voicemail Detection

Those speed and accuracy trade-offs play out differently depending on the underlying infrastructure. Most AI voice platforms route calls through a third-party CPaaS (Communications Platform as a Service) like Twilio, which means AMD runs on rented infrastructure and caller ID is not issued at the carrier level. Plura AI’s AI Predictive Dialer runs on Plura’s own FCC-licensed carrier stack, which changes what is possible at the infrastructure level.

Key differentiators for outbound teams evaluating AMD at scale:

• FCC-licensed carrier infrastructure: AMD runs on Plura’s own stack, not a third-party CPaaS wrapper. Voice originates on domestic infrastructure with STIR/SHAKEN authentication on every outbound call.
• Branded caller ID at the carrier level: Calls present with the company name rather than an unknown number, which reduces “Spam Likely” labels and improves answer rates before AMD even runs. Plura’s integration with Number Verifier improves outbound call connection rates by up to 45% through built-in spam prevention.
• Compliance infrastructure built in: Real-time DNC scrubbing, TCPA-litigator screening, and quiet-hours enforcement run before every dial. Consent records are timestamped and immutable. Plura supports compliance with TCPA, DNC requirements, HIPAA-related safeguards, SOC 2, ISO certification programs, GDPR frameworks, and STIR/SHAKEN caller ID verification.¹ Customers remain responsible for their own regulatory obligations and certifications.
• Stateful conversation memory: The same platform that handles voicemail detection on outbound calls shares conversation context across voice, SMS, RCS, and webchat. A lead that went to voicemail on one channel can be followed up on another with full prior context.

For teams comparing Plura vs. Bland AI or evaluating plans and rates, the carrier-level distinction is the operational difference that matters most at scale.

Voicemail detection accuracy is a measurable, tunable metric. The published range runs from 60% for legacy heuristics to 98.5% for fine-tuned AI models, but the number only matters alongside detection window, false positive rate, and production conditions. For high-volume outbound teams, Plura’s AI Predictive Dialer delivers AMD on an FCC-licensed carrier stack with branded caller ID, real-time DNC scrubbing, and TCPA compliance support built in before every dial.

To see carrier-grade AMD in production, compare plans and rates side by side, then book a live demo with Plura.

---

Read Next

• How to Improve Voicemail Detection Accuracy: 2026 Guide
• How Accurate Is Voicemail Detection? 2026 Benchmarks
• Voicemail Detection Accuracy Benchmarks for Outbound Teams
• Voicemail Detection Best Practices for AI Predictive Dialers
• Voicemail Detection Integration for Contact Centers