How to Reduce AMD False Positives in Your Contact Center

Written by: Matt Beucler, CEO, Plura AI

Why This Guide Matters

This guide explains why voicemail detection false positives happen, how to measure and tune your AMD settings, and when AI-based detection makes financial sense.

Voicemail Detection Errors Quietly Drain Revenue and Strain Compliance

When AMD misclassifies a live human as a machine, the system drops the call or routes it into voicemail-drop logic. The prospect hears dead air and a hang-up. The lead is burned and the agent sits idle. Because machine-classified calls are excluded from the FCC’s 3% abandoned-call cap under 47 CFR 64.1200(a)(7), reported compliance metrics can look healthy while real consumers are being disconnected.¹ Regulators and plaintiffs’ attorneys focus on this pattern when a TCPA complaint triggers a call log review.

The revenue impact is direct and measurable. A 20,000-dial-per-day operation with a 12% connect rate and a 20% false positive rate drops 480 live calls daily, which costs $19,200 per day at $40 per conversation. Reducing that rate to 3% recovers over $490,000 per month.³ ViciStack’s testing across 2.4 million calls in Q1 2026 found that false positive rates vary by carrier, from 3.4% on AT&T landlines to 11.3% on Google Voice.³

Every false positive represents a live human conversation that never happens. The problem keeps growing as mobile carriers, iOS 26 call screening, and VoIP audio processing reshape the acoustic signals that legacy AMD systems were calibrated on decades ago.

See Plura AI’s AI Predictive Dialer in a live demo to understand how AI-based dialing reduces false positives and increases agent talk time.

Root Causes of Voicemail Detection False Positives

Legacy AMD systems like Asterisk’s app_amd use rules-based silence detection and word counting rather than machine learning.⁴ They analyze the first few seconds of audio against fixed thresholds. When a human’s speech pattern resembles a voicemail greeting, the system misclassifies the call. Asterisk’s AMD application processes 20-millisecond audio frames at 8 kHz and acts as a silence detector with a word counter.

Common root causes of false positives include:

• Long or formal greetings: A person answering “Hello, thanks for calling [Company], this is Sarah speaking” trips word-count and greeting-length thresholds that target machine greetings.
• Initial silence and pauses: Mobile users often pause before speaking. If initial_silence is set too low at the 2500 ms default, the system assumes a machine greeting has started.
• Background noise: Office murmur, TV audio, or traffic can resemble the acoustic profile of a recorded message.
• Aggressive timeouts: Setting total_analysis_time too low forces the algorithm to guess under time pressure.
• Carrier audio processing: Modern VoIP carriers apply noise reduction and level normalization that distort the audio signatures legacy AMD relied on.
• iOS 26 voicemail greetings: Apple’s updated voicemail greetings are shorter and acoustically different, which confuses threshold-based detection.

These failure modes appear in Asterisk’s AMD configuration documentation and Twilio’s AMD documentation. VICIdial’s creator Matt Florell told an Astricon audience in 2012 that his team “almost never suggests using” stock AMD.

How to Measure Your AMD False Positive Rate

Accurate tuning starts with a baseline measurement of your actual false positive rate. Without that baseline, you cannot tell whether a change improves or degrades performance.

Use these diagnostic prompts to frame your review:

• Check whether false positives occur more on mobile numbers than landlines.
• Watch for calls that drop during the first two seconds after pickup.
• Document your current AMD sensitivity settings per campaign.
• Compare MACHINE classification rates across carriers.

Use this measurement methodology, based on amdify.io’s audit process:

1. Enable call recording for AMD-classified calls.
2. Pull a random sample of 200 calls classified as MACHINE.
3. Listen to each recording and count calls with audible human responses.
4. If 30 of 200 have human responses, your false positive rate is approximately 15%. If 50 do, it is 25%.

An alternative A/B method routes 10-20% of campaign traffic with AMD disabled. Agents manually code outcomes. You then compare the MACHINE rate from manual coding versus AMD during the same period. Review the AMD Log Report in VICIdial to identify patterns by carrier, geography, and campaign.

Platform-Specific Tuning to Reduce False Positives

AMD tuning always balances false positives against false negatives. Looser parameters protect live humans but send more actual voicemails to agents. The practical goal is a configuration that keeps false positives low while keeping voicemail volume manageable.

VICIdial and Asterisk tuning parameters:

Per ViciStack’s production baseline and amdify.io’s false positive reduction guide, edit /etc/asterisk/amd.conf with these changes:

Parameter	Default	False Positive Reduction	Purpose
initial_silence	2500 ms	3000 ms	Gives more time before assuming machine
greeting	1500 ms	2000 ms	Allows longer human greetings
maximum_number_of_words	2	4-5	Allows humans to say more before machine classification
min_word_length	100 ms	80 ms	Catches shorter human responses like “Hello?”
total_analysis_time	5000 ms	5500 ms	Gives the algorithm more time to decide

Change one parameter at a time, dial a few hundred leads, and pull the AMD Log Report before making another change. This isolates the effect of each adjustment. Tune per campaign rather than globally, because cell-heavy lists behave differently than landline lists.

Twilio AMD settings:

Twilio’s AMD uses machine learning and exposes sensitivity parameters.⁴ Set machineDetection to DetectMessageEnd for a conservative approach. Increase the machineDetectionSpeechThreshold and machineDetectionSpeechEndThreshold values to reduce false positives on hesitant speakers. A higher speech threshold requires more speech before classifying a call as human.

AWS Connect:

Use the AnsweringMachineDetection contact attribute with a conservative configuration.⁴ Set machineDetectionTimeout to at least 4000 ms. Configure machineDetectionSpeechThreshold to require more speech before human classification.

False positive rates vary significantly by carrier. T-Mobile’s short, casual voicemail greetings produce an 8.1% false positive rate even when tuned, while Google Voice’s screening feature hits 11.3%. When one or two carriers drive most of your errors, tuning alone rarely solves the issue.

Managing the False Positive vs False Negative Trade-Off

Every AMD tuning decision trades false positives against false negatives. A false positive, where a human is classified as a machine, wastes a live lead. The prospect experiences a dropped call and the conversation never happens. A false negative, where a machine is classified as human, connects an agent to voicemail and wastes 15-30 seconds of talk time before manual dispositioning.

The asymmetry matters: an agent can disposition a voicemail in about 5 seconds, while a lost live connection is gone permanently. For sales campaigns where every conversation has revenue potential, bias settings toward false negatives and let more machines through to protect live humans. For survey or informational campaigns where talk time is the main constraint, bias settings toward false positives.

ViciStack recommends a target of 85% or higher accuracy with a false positive rate under 5% as an operational benchmark for voicemail drop campaigns.

Watch a live demo of Plura AI’s AI Predictive Dialer to see how it manages this trade-off automatically.

When AI-Based AMD Becomes the Better Tool

Legacy AMD reaches a structural accuracy ceiling. Most operations that tune aggressively reach an 8-12% false positive rate. Below that level, the remaining error is structural and does not respond to parameter changes. ViciStack’s testing across more than 100 VICIdial deployments found that even per-carrier tuned traditional AMD achieves only 85-92% accuracy with a 4-8% false positive rate.

AI-based AMD uses machine learning to analyze temporal audio patterns and speech dynamics. ViciStack’s benchmarks show AI AMD reaches 92-96% overall accuracy with a 2-4% false positive rate.³

Criterion	Legacy AMD (Tuned)	AI-Based AMD
Overall accuracy	85-92%	92-96%
False positive rate	4-8% tuned; 15-25% default	2-4%
Detection method	Rules-based silence and word counting	Machine learning on speech patterns
Adaptability	Fixed thresholds with manual tuning	Continuous learning from call audio
Carrier variation (false positive rate)	3.4-11.3% by carrier	1.9-4.7% by carrier

Operations running more than 300 dials per hour should evaluate AI-powered detection as a replacement for Asterisk AMD. If tuning cannot get your false positive rate below roughly 8-10%, rules-based AMD has reached its structural ceiling.

Plura AI’s AI Predictive Dialer is one such AI-based option. It connects agents only to live humans, which removes voicemail detection false positives at the source. As an FCC-licensed carrier, Plura issues branded caller ID and runs STIR/SHAKEN authentication on every outbound call.¹ These capabilities support higher connect rates and reduce the “Spam Likely” labeling that suppresses answer rates. The same AI Predictive Dialer uses stateful conversion signals, including historical answer rates and prior negotiation outcomes, to prioritize the contacts most likely to convert.

Keeping AMD Performance Stable Over Time

AMD performance changes over time. Carrier routing changes, mobile OS updates, list composition shifts, and seasonal traffic patterns all affect accuracy. That is why you should monitor AMD performance by carrier, geography, and campaign. Review the AMD Log Report weekly, targeting less than 5% NOTSURE-TOOLONG and less than 2% NOTSURE-NOAUDIODATA.

Operational practices that sustain accuracy over time:

• Run A/B tests when changing parameters and wait at least 48 hours before evaluating results.
• Tune per campaign rather than globally. Cell-heavy lists and landline lists require different configurations.
• Segment accuracy analysis by carrier. Carrier-specific false positive rates vary by as much as 7.9 percentage points on traditional AMD.
• Re-run the tuning process from baseline whenever iOS updates, carrier routing changes, or list composition shifts significantly.
• If a market or list is consistently noisy, consider disabling AMD for that campaign and handling voicemail manually.

Frequently Asked Questions

What causes AMD false positives?

AMD false positives occur when rules-based detection misclassifies a live human as a machine. Legacy systems like Asterisk’s app_amd focus on timing and word count rather than speech content. Common triggers include long or formal greetings that exceed word-count thresholds and initial silence or pauses that mimic a voicemail lead-in. Background noise can also resemble a recorded message. Aggressive timeout settings force the algorithm to guess before it has enough audio, and carrier audio processing distorts the acoustic signatures the system was originally calibrated on. iOS 26 introduced shorter, acoustically distinct voicemail greetings that further confuse threshold-based detection. Parameters like initial_silence, greeting, and maximum_number_of_words set too aggressively are the most common configuration-level causes.

How do I reduce AMD false positives in VICIdial?

Edit /etc/asterisk/amd.conf with these changes: increase initial_silence to 3000 ms, increase maximum_number_of_words to 4-5, and decrease min_word_length to 80 ms. Increasing the greeting parameter to 2000 ms also helps by allowing longer human greetings before the system commits to a machine classification. Change one parameter at a time, dial a few hundred leads, and review the AMD Log Report before making another change. This approach keeps each adjustment isolated. Tune per campaign rather than applying a single global configuration. Segment your accuracy analysis by carrier, since T-Mobile and Google Voice produce significantly higher false positive rates than AT&T landlines even with identical settings. After each change, wait at least 48 hours to collect a statistically meaningful sample before evaluating results.

What is a good AMD false positive rate?

A false positive rate under 5% with at least 85% overall accuracy is a practical benchmark for voicemail drop campaigns. Default Asterisk parameters typically produce 15-25% false positives on mixed traffic. Manual tuning can reach the 8-12% structural ceiling described earlier for most operations. Per-carrier tuned traditional AMD can reach 4-8%. AI-based AMD typically operates below that ceiling and achieves lower false positive rates within the accuracy ranges discussed earlier. When your tuned false positive rate stays above 10%, parameter changes alone rarely resolve the issue. At rates above 15%, the operation leaves meaningful revenue on the table every day.

How much do AMD false positives cost?

The cost depends on call volume, connect rate, and the value of each conversation. As shown earlier, a high-volume operation can lose more than $490,000 per month to false positives. At a 50-agent center running 3,000 live answers per day, a 20% false positive rate results in 600 dropped live calls daily. Beyond direct revenue loss, false positives also damage caller ID reputation when prospects hear dead air and flag the number, which can suppress answer rates across the entire campaign over time.

When should I upgrade to AI-based AMD?

Operations running more than 300 dials per hour should evaluate AI-powered detection as a replacement for legacy AMD. The clearest signal is a false positive rate that cannot be tuned below the 8-10% structural floor despite systematic parameter adjustment. Rules-based AMD cannot distinguish between a human saying “Hey, it is Sarah, what is up?” and a voicemail greeting that says “Hey, it is Sarah, leave a message” because it has no semantic understanding of speech content. AI-based AMD analyzes temporal audio patterns and speech dynamics and reaches the accuracy ranges discussed earlier. For B2C cell-phone-heavy campaigns, the accuracy gap between tuned traditional AMD and AI AMD often reaches 8-13 percentage points. For B2B office-line campaigns, the gap narrows to 2-4 percentage points, which may not justify the additional cost and latency for smaller operations.

Stop Dropping Live Humans

Voicemail detection false positives are manageable when you measure your actual rate, tune systematically, and recognize when legacy AMD has reached its structural ceiling. High-volume operations gain the most by moving to AI-based AMD that connects agents only to live humans. Plura’s AI Predictive Dialer supports this approach and runs on Plura’s FCC-licensed carrier with branded caller ID and STIR/SHAKEN authentication. Compare plans and rates.

Schedule a live walkthrough of Plura AI’s AI Predictive Dialer to see how AI-powered dialing reduces false positives and increases productive talk time.

---

Read Next

• How to Improve Voicemail Detection Accuracy: 2026 Guide
• Voicemail Detection Integration for Contact Centers
• Predictive Dialer Voicemail Detection: How AMD Works
• Voicemail Detection Best Practices for AI Predictive Dialers
• How Voicemail Detection Improves Agent Utilization