Vonage AMD Accuracy: What the Numbers Actually Mean

Written by: Matt Beucler, CEO, Plura AI

Key Takeaways

• Vonage does not publish a current production accuracy figure for Advanced Machine Detection (AMD), so operators must measure performance against their own call recordings.
• AMD accuracy depends on a structural latency tradeoff. The system must classify human versus machine inside a fixed 3–5 second audio window before the full greeting plays.
• The 60–80% accuracy range cited in developer discussions reflects a typical industry range for real-time acoustic analysis, not a Vonage production guarantee.³
• Operators can calculate meaningful false-positive and false-negative rates by running a confusion-matrix procedure on 1,000–2,000 labeled calls from a single stable campaign.
• Teams that need higher accuracy and carrier-grade controls can see how AI-powered voicemail detection performs at scale with Plura AI.

The Problem: Why Vonage AMD Accuracy Is Hard to Pin Down

Operators who enable AMD on Vonage’s Voice API see misclassifications in production and then look for a Vonage benchmark to compare against.⁴ That benchmark does not exist because Vonage does not publish a current overall production accuracy figure for Advanced Machine Detection.

The reason is structural. AMD must render a verdict before the full greeting is available. The system commits to a human-or-machine classification inside a fixed audio window. The verdict is therefore a timing tradeoff rather than a classification certainty. Because that tradeoff is unavoidable, every misclassification carries a direct operational cost. A false negative leaves dead air on a voicemail system while an agent waits. A false positive drops a live prospect before any conversation begins.

Test carrier-grade voicemail detection against your own call population with a live Plura demo.

How Does Vonage AMD Decide Human vs. Machine?

To understand why accuracy is hard to pin down, it helps to see how the detection actually works. Vonage’s Voice API Advanced Machine Detection analyzes the call audio once the call reaches the answered state and classifies the answer as human or machine, with machine events carrying a sub_state of beep_start for voicemail beeps, fax for fax machines, or beep_timeout when no beep is received. This classification happens synchronously, meaning the call waits for detection before fetching the XML response, which adds 2–5 seconds of silence at the start of the call.

The detection mechanism analyzes the tone of voice and the speed of the words spoken in the greeting rather than looking for a single definitive signal. Vonage’s Advanced Machine Detection exposes three modes beyond the standard offering: default, detect, and detect_beep. The beep_start sub_state fires when the system identifies the characteristic beep tone that follows a voicemail greeting, which signals that the caller’s audio channel is ready to record a message.

The model must commit before the full greeting is available. That constraint is a structural consequence of low-latency audio classification, not a Vonage-specific limitation. Research on real-time turn detection identifies a fundamental latency-accuracy tradeoff. Systems that wait for higher-quality semantic information are more accurate but slower, while faster systems sacrifice accuracy. When a greeting is too short to yield confident classification, any streaming detector faces the same ambiguity.

Why Does Vonage Report 60–80% Accuracy?

The 60–80% range represents a category baseline for real-time acoustic analysis, not a Vonage production guarantee.³ A 2026 paper by Kumar Saurav reports that industry-standard AMD based on silence heuristics achieves only 70–85% accuracy with 10–15% false-positive rates.³ A June 2026 DialerBee analysis similarly reports that predictive dialers achieve around 75% accuracy and legacy rule-based detection can fall to around 40% in difficult conditions.³

The 96% figure that circulates in developer discussions does not describe Vonage’s production AMD service. It comes from a 2026 arXiv paper by Kumar Saurav titled “Real-Time Voicemail Detection in Telephony Audio Using Temporal Speech Activity Features” (arXiv:2604.09675), which describes a standalone machine-learning model evaluated on 764 telephony recordings. That model achieved 96.1% combined accuracy using 15 temporal features from a neural voice activity detector (VAD) and a shallow tree-based ensemble. It describes a separate research implementation. Vonage’s production AMD is a callback-driven call-answer classification feature exposed through the Voice API, not a published machine-learning research model, and its internal implementation is not presented as the same project behind the 96% figure.

Standard vs. Advanced Machine Detection: What Features Change?

Vonage offers two tiers of AMD. Standard AMD provides basic detection. Advanced Machine Detection adds features including beep detection and finer call-control options (asynchronous mode) and is a chargeable premium feature with an extra charge of €0.0070 per call on which Advanced Machine Detection is activated. Within Advanced Machine Detection, three modes target different operational needs.

Mode	What It Adds	Source
Standard AMD	Basic human/machine classification on call audio	API Voice AMD documentation
Advanced: default	Improved classification with webhook-delivered verdict	API Voice AMD documentation
Advanced: detect	Same behavior as Standard Machine Detection with more accurate machine identification	API Voice AMD documentation
Advanced: detect_beep	Fires beep_start sub_state when beep tone is detected and enables message drop after beep	API Voice AMD documentation

The Vonage AMD versus detect_beep mode comparison is the most operationally significant configuration decision for voicemail campaigns. detect_beep reduces false negatives on systems that emit a standard beep tone. However, only 0.3% of voicemail recordings contain a detectable beep on the callee’s audio channel, which limits the mode’s coverage in production. Operators running campaigns against populations with no-beep voicemail systems will see limited benefit from detect_beep alone.

Why Vonage AMD Often Fails on Google Voice Numbers

Google Voice screening is a known failure case for voicemail detectors because Google Voice asks callers to identify themselves, which sounds like a live conversation rather than a voicemail prompt. Short initial greetings reduce the amount of acoustic evidence available in the first seconds of a call, which makes real-time detectors more likely to classify them as human.

This pattern is a structural consequence of the evidence bottleneck in low-latency audio classification, not an isolated Vonage bug. Voice activity detection methods are described in the research literature as “semantically blind.” They rely on audio energy levels and cannot distinguish between a speaker finishing a sentence and pausing for breath or thought. When a greeting is too short to yield confident classification, any streaming detector faces the same ambiguity. The model must commit before enough signal is available. Some short-greeting false positives sit at the inherent boundary of synchronous acoustic detection rather than a tunable parameter.

How Do You Measure Vonage AMD Accuracy Yourself?

Given these structural limitations, the only way to know how AMD performs on your traffic is to measure it directly. The procedure below produces a confusion matrix against your own labeled recordings. Run it on a single stable campaign to isolate the variables. A minimum of 1,000–2,000 dialed calls from a single stable campaign is recommended to get a statistically meaningful false-positive and false-negative read, because smaller samples swing several points due to daily answer-pattern variance.

1. Capture and label recordings from a single stable campaign. Pull a two-week dataset from one campaign with consistent dialing parameters, lead source, and AMD configuration. Export call recordings and AMD disposition codes (human, machine, unknown) for the full sample.
2. Define ground truth by manually re-scoring a sample of machine-classified calls. Manually re-score 150–200 calls that AMD classified as machine by listening to recordings and marking each as either a correct machine detection or a false positive where a live person answered. Ground truth is the human-labeled verdict, not the AMD verdict.
3. Compute your false-positive rate (FPR). FPR is the share of calls where a live human answered but AMD classified the call as a machine, which causes the call to be dropped or routed to a voicemail-drop flow. Divide the count of false positives by the total count of machine-classified calls in your re-scored sample.
4. Compute your false-negative rate (FNR). FNR is the share of calls where a voicemail system answered but AMD classified the call as human, routing the call to a live agent who waited through a voicemail greeting. Compute this from a parallel sample of human-classified calls re-scored against recordings.
5. Track both rates by greeting type. Segment results by short greetings (under 3 seconds), standard carrier voicemail greetings, custom greetings, and automated attendants. Varying greeting length deliberately in testing at three seconds, ten seconds, and twenty-five seconds is recommended because each behaves like a different detection problem.
6. Re-audit AMD accuracy on a regular cadence. Weekly aggregate machine versus human split checks by campaign, monthly manual re-scores of 150–200 calls per active campaign, and quarterly full carrier and lead-source segmentation are recommended, with an immediate 50–100 call spot-check after any configuration or carrier change.

Common Vonage AMD Failure Patterns and Mitigations

The following patterns account for most AMD misclassifications in production Vonage deployments.

• Long greetings from business receptionists. A greeting like “Hello, this is Sarah at Coastal Realty, how can I help you?” reads as a machine because it is long and declarative. Mitigation: increase the speechThreshold parameter to 3,000–4,000 ms so the system waits for a longer continuous speech burst before classifying as machine.
• No-beep voicemail systems. Some voicemail systems use silence or verbal prompts like “Please leave a message” instead of a beep. detect_beep mode provides no benefit on these systems. Mitigation: use detect mode rather than detect_beep for campaigns targeting populations with diverse voicemail implementations.
• Automated attendants that sound human. Business auto-attendants that answer with a menu and slow answerers who pause before speaking are additional edge cases that confuse voicemail detectors. Mitigation: route uncertain classifications to a live agent rather than dropping the call.
• Silent human answers. A human who picks up and says nothing for the first second shifts the AMD timer and can trigger a machine classification. Mitigation: shorten the initial silence threshold and bias uncertain results toward human.
• Network audio quality. Compressed codecs such as G.729 discard waveform detail and raise AMD false-positive rates. Native G.711u pass-through at the full 64 kbps is the standard codec recommendation for dialer trunks. Mitigation: verify codec configuration on your SIP trunk and carrier route.

When to Stop Tuning Vonage AMD and Route Around It

Some operators eventually reach the ceiling of synchronous AMD tuning. A 15–25% false-positive rate often reflects the ceiling of energy and duration-based AMD detection, which is why the industry has moved toward AI classification models trained on greeting audio. When the confusion-matrix procedure shows that FPR remains above 10% after parameter tuning, the synchronous model’s architecture usually acts as the constraint rather than the configuration.

Plura AI is built for operators at that ceiling. Plura is its own FCC-licensed audio bridging carrier, which means voice does not route through a third-party CPaaS (Communications Platform as a Service). Voicemail detection and the AI Predictive Dialer run on that same carrier foundation. This carrier ownership enables branded caller ID issued at the carrier level and STIR/SHAKEN (Secure Telephone Identity Revisited/Signature-based Handling of Asserted information using toKENs) authentication on every outbound call. Additionally, the platform supports compliance before dial with real-time DNC (Do Not Call) scrubbing, TCPA (Telephone Consumer Protection Act) litigator screening, automated quiet hours, and immutable consent logging.^1,2 The AI Predictive Dialer uses stateful conversion signals to decide who to call next, rather than relying on static list order.

Explore how the carrier-grade dialer handles voicemail detection at scale in a live Plura demo.

Frequently Asked Questions

How Accurate Is Vonage Voicemail Detection?

Vonage does not publish a current overall production accuracy figure for Advanced Machine Detection. The 60–80% range discussed earlier reflects a typical category range for real-time acoustic analysis based on energy and duration heuristics. Modern AI-based voicemail detection systems trained on telephony audio often report higher accuracy figures, but each vendor measures those results on its own dataset. Those numbers do not transfer directly to a Vonage deployment with different carriers, greetings, and lead populations. The only accuracy number that applies to your operation is the one you compute from your own labeled recordings using the confusion-matrix procedure above.

What Is the Difference Between Vonage AMD and Advanced Machine Detection?

Standard AMD is Vonage’s basic answering machine detection capability. Vonage’s Advanced Machine Detection, a chargeable premium feature (€0.0070 per call), adds beep detection with the beep_start webhook sub_state in detect_beep mode, asynchronous mode for better call control, and configurable behavior, beep_timeout, and mode parameters. The three Advanced modes (default, detect, detect_beep) each target a different error pattern. default improves overall classification. detect provides more accurate machine identification for better routing. detect_beep fires a specific sub_state when a beep tone is detected so a message can be dropped precisely after the beep. The right mode depends on which error type costs your operation more: false positives that drop live prospects or false negatives that keep agents on voicemail greetings.

How Do I Measure My Own Vonage AMD Accuracy?

Pull 1,000–2,000 calls from a single stable campaign. Manually re-score 150–200 calls that AMD classified as machine by listening to recordings and marking each as a correct machine detection or a false positive. Compute your false-positive rate by dividing confirmed false positives by the total machine-classified sample. Run a parallel re-score on human-classified calls to compute your false-negative rate. Segment both rates by greeting type: short greetings under 3 seconds, standard carrier voicemail greetings, custom greetings, and automated attendants. Re-audit monthly and run a spot-check of 50–100 calls after any configuration or carrier change.

Why Does Vonage AMD Misclassify Google Voice Numbers?

Google Voice asks callers to identify themselves before connecting the call. That screener prompt sounds like a live conversation rather than a voicemail greeting, so AMD often classifies it as human. The underlying cause is the evidence bottleneck in low-latency audio classification. When a greeting is too short or too interactive to yield confident classification inside the 3–5 second detection window, any streaming detector faces the same ambiguity. This pattern reflects a structural consequence of synchronous acoustic detection rather than a Vonage-specific bug. Similar failure patterns appear across carrier-level AMD implementations from multiple providers.

Does the 96% Accuracy Figure Apply to Vonage Production AMD?

The 96% figure comes from a 2026 arXiv paper by Kumar Saurav (arXiv:2604.09675) describing a standalone machine-learning model evaluated on 764 telephony recordings. That model uses 15 temporal features from a neural voice activity detector and a shallow boosted tree ensemble, achieving 96.1% combined accuracy with 46 ms inference latency on a commodity dual-core CPU. It describes a separate research implementation rather than Vonage’s production AMD service. Vonage’s production AMD is a productized API capability whose internal implementation is not presented as the same research project.

Conclusion: The Only Accuracy Number That Matters Is Yours

Vonage AMD accuracy functions as an operational metric that you measure against your own call population, lead sources, carrier routes, and greeting mix. The 60–80% category baseline describes the typical accuracy range for real-time acoustic analysis, not a Vonage-published production guarantee. The 96% figure describes a separate research model. Neither number tells you what is happening on your campaigns today.

Run the confusion-matrix procedure against your own labeled recordings. Compute FPR and FNR by greeting type. If tuning the Advanced Machine Detection mode parameters closes the gap, you can continue with the synchronous model. If false-positive rates remain elevated after tuning, it may be worth evaluating a carrier-grade platform with voicemail detection engineered into the dialer layer.

For teams comparing options, the evaluation criteria that matter include carrier ownership with branded caller ID issued at the carrier level, compliance controls enforced before dial such as real-time DNC scrubbing and TCPA litigator screening, and stateful conversation memory across every channel.

Compare plans and rates side by side.

Run your numbers through Plura’s calculator to check your ROI in real time.³

---

Read Next

• How to Improve Voicemail Detection Accuracy: 2026 Guide
• How Accurate Is Voicemail Detection? 2026 Benchmarks
• Voicemail Detection Accuracy: 2026 Benchmarks and Tuning
• Voicemail Detection Best Practices: 2026 Tuning Guide
• Voicemail Detection Accuracy Benchmarks for Outbound Teams