AI Answering Machine Detection
SpeechLLM detects human vs. voicemail in under 50ms, from the first audio frame — 99.7% accuracy across 50M+ analyzed calls.
No credit card required
The problem
Legacy AMD hangs up on 15-20% of real people
Rule-based answering machine detection guesses from timing and silence thresholds alone, producing a 15-20% false positive rate. Every false positive is a live prospect dropped mid-greeting and an agent left waiting on a dead line.
VM Hunter replaces that guesswork with 8 signal layers analyzed in parallel, cutting the false positive rate to 0.2% at 99.7% overall accuracy.
Legacy AMD
15-20%
False positive rate — live humans dropped
VM Hunter
0.2%
False positive rate — 99.7% accuracy overall
How it works
SpeechLLM analyzes 8 signal layers in parallel
Transcription-only systems wait for words before they can decide. SpeechLLM evaluates all 8 layers simultaneously from the first audio frame, so a classification is available before a greeting finishes.
- 01
Beep tones & SIT signals
Catches special information tones that mark disconnected or invalid numbers.
- 02
DTMF & fax tone recognition
Identifies keypad tones and fax handshakes that never belong to a live prospect.
- 03
Energy level profiling
Measures per-frame amplitude to separate speech from recorded playback levels.
- 04
Dead air tracking
Times silence windows that indicate a greeting has ended and recording has begun.
- 05
Static & noise fingerprinting
Profiles carrier noise and codec artifacts so compression never skews the result.
- 06
Pre-recorded audio detection
Flags the acoustic signature of playback audio versus a live open microphone.
- 07
Live transcription
Streams speech-to-text so the words themselves become a classification signal.
- 08
LLM phrase classification
Resolves ambiguous cases by reading greeting phrasing for intent and meaning.
The output payload
Results are delivered over WebSocket as 8kHz PCM audio in 20ms chunks with zero buffering delay.
- AMDSTATUS
- HUMAN / MACHINE / CALLGUARD / DISCONNECT
- AMDCAUSE
- The specific layer and signal that drove the decision
- CONFIDENCE
- Per-classification score for custom routing thresholds
99.7%
Detection accuracy
<50ms
Response time
10,000+
Concurrent calls supported
99.99%
Uptime SLA
50+
Languages supported
50M+
Call recordings trained on
Results
Agent utilization from 40% to 85%
Every voicemail an agent listens to is paid time producing zero revenue. After switching, customers move agent utilization from 40% to 85% — more than doubling the share of the day spent in live conversations without adding a single seat.
Because the gain comes from recovered talk time rather than new headcount, ROI is typically visible within the first week of deployment.
Agent utilization
+45 percentage points of recovered selling time per agent, per shift.
Integration
Live in about 30 minutes
A REST API plus SDKs for all major languages, and native VICIdial compatibility using standard AMD variables — no dialplan rewrites and no infrastructure replacement.
Read the integration docsSecurity
- SOC 2 Type II compliant
- End-to-end encryption in transit and at rest
- Call data never leaves our secure infrastructure
const socket = new WebSocket("wss://api.vmhunter.com/v1/stream", {
headers: { Authorization: "Bearer YOUR_API_KEY" },
})
// Stream 8kHz PCM audio in 20ms chunks
socket.send(pcmChunk)
socket.onmessage = (event) => {
const { AMDSTATUS, AMDCAUSE, CONFIDENCE } = JSON.parse(event.data)
// AMDSTATUS: "HUMAN" | "MACHINE" | "CALLGUARD" | "DISCONNECT"
if (AMDSTATUS === "HUMAN") routeToAgent()
}Comparison
VM Hunter vs. legacy AMD
| Criterion | Legacy AMD | VM Hunter |
|---|---|---|
| Detection method | Single-signal timing rules or transcription only | 8 signal layers analyzed in parallel by SpeechLLM |
| False positive rate | 15-20% of live humans classified as machines | 0.2% false positives at 99.7% overall accuracy |
| Response time | 2-5 seconds of greeting required before deciding | Under 50ms from the first audio frame |
| Language support | English-only, degrades on regional accents | 50+ languages with accent-aware models |
| Setup time | Days of dialplan tuning and threshold tweaking | About 30 minutes with standard AMD variables |
Pricing
Start free with 5,000 calls per month
The free plan includes 5,000 calls per month with full access to core detection features and no credit card required.
Paid tiers scale for professional and enterprise call centers, with savings on annual billing.
Compare plansFAQ
Frequently asked questions
What is AI Answering Machine Detection?
AI Answering Machine Detection uses machine learning to decide, in real time, whether an outbound call was answered by a live person or a voicemail system. Instead of relying on timing and silence thresholds like rule-based AMD, VM Hunter's SpeechLLM analyzes 8 signal layers in parallel from the first audio frame. It returns an AMDSTATUS of HUMAN, MACHINE, CALLGUARD, or DISCONNECT with a confidence score in under 50ms.
How accurate is VM Hunter's AI Answering Machine Detection?
VM Hunter delivers 99.7% detection accuracy, measured across more than 50 million analyzed calls in production. The false positive rate is 0.2%, compared with the 15-20% typical of legacy rule-based AMD. Accuracy holds across 50+ languages and regional accents.
How is AI AMD different from traditional Asterisk or VICIdial AMD?
Traditional Asterisk and VICIdial AMD apply fixed timing heuristics — greeting length, silence duration, and word count — which forces a tradeoff between dropping live humans and sending voicemails to agents. VM Hunter evaluates tones, energy levels, dead air, noise fingerprints, pre-recorded audio signatures, live transcription, and LLM phrase classification simultaneously. It sets the same AMDSTATUS and AMDCAUSE variables your dialplan already reads, so no rewrite is required.
How fast is the detection response time?
Classification returns in under 50ms, measured from the first audio frame rather than the end of the greeting. Audio streams over WebSocket as 8kHz PCM in 20ms chunks with zero buffering delay. The decision is available before most greetings have finished playing.
Does AI AMD work with VICIdial?
Yes. VM Hunter is VICIdial compatible and sets the standard AMDSTATUS, AMDCAUSE, and confidence variables that VICIdial dialplans already use. Most teams are live in about 30 minutes without replacing any infrastructure.
Is VM Hunter's AMD SOC 2 compliant?
VM Hunter is SOC 2 Type II compliant. Call audio is encrypted end to end, in transit and at rest, and never leaves our secure infrastructure. The platform also carries a 99.99% uptime SLA.
How long does integration take?
Most customers are live in about 30 minutes using the REST API or one of the SDKs available for all major languages. Because VM Hunter reuses standard AMD variables, there are no dialplan rewrites. Full quickstarts and VICIdial guides are in the docs.
Stop paying agents to listen to voicemail
Deploy 99.7% accurate detection in about 30 minutes and see recovered talk time in your first week.
Or compare pricing plans first.