AMD: Answering Machine Detection Engine

In outbound calling, 30-40% of calls reach voicemail. Every second an agent waits for a machine to finish its greeting is wasted labor. The AMD Engine classifies answering entities in under 3 seconds using a four-feature decision tree — detecting humans, voicemail systems, beep tones, and unknown entities with explainable confidence scores.

1. The Outbound Calling Problem

Outbound contact centers face a persistent efficiency challenge: a significant fraction of calls connect to answering machines rather than humans. Without AMD:

• Agents wait 5-15 seconds listening to voicemail greetings before realizing no human is present
• False positives (human classified as machine) cause premature hangups and lost customers
• False negatives (machine classified as human) waste agent time on voicemail recordings
• Beep detection failures mean agents miss the opportunity to leave automated messages

Traditional AMD systems rely on single features (usually silence duration or greeting length) and take 5-8 seconds to reach a verdict — by which time, a human caller has already said “hello” and heard silence.

2. Four-Feature Decision Architecture

The AMD Engine extracts four orthogonal features from the first 3 seconds of audio:

3. Decision Tree Classifier

The four features feed into a priority-ordered decision tree:

Why a Decision Tree?

The AMD classifier deliberately uses a rule-based decision tree rather than a neural network for three reasons:

1. Explainability — every verdict comes with a human-readable reason (“initial_silence=1800ms exceeds human threshold of 1500ms”)
2. Calibration — thresholds can be tuned per carrier, region, or campaign without retraining
3. Determinism — the same audio always produces the same verdict, critical for audit compliance

4. AMD Contracts

@dataclass(frozen=True)
class AmdFeatures:
    initial_silence_ms: int       # Silence after connect
    greeting_duration_ms: int     # First speech segment length
    speech_continuity: float      # 0.0-1.0 flow score
    cadence_regularity: float     # 0.0-1.0 rhythm uniformity
    beep_tone_detected: bool      # Frequency-domain beep flag
    beep_frequency_hz: float      # Dominant beep frequency
    background_noise_level: float # Ambient noise floor
    total_duration_ms: int        # Total analyzed duration

@dataclass(frozen=True)
class AmdResult:
    result_id: str
    correlation_id: str
    ts_ms: int
    verdict: str       # "human" | "machine" | "beep" | "unknown"
    confidence: float  # 0.0 - 1.0
    features: AmdFeatures
    reason: str        # Human-readable explanation

5. SignalOS Integration

AMD results automatically convert to SignalOS signals:

Downstream Actions

6. Evidence Pack Integration

AMD results are automatically included in the session evidence pack:

• AMD Card: verdict, confidence, all four feature values, decision reason
• Timeline: AMD classification timestamp relative to call connect
• Audit Trail: feature thresholds used, calibration version, carrier info

This ensures that every AMD decision is traceable and auditable for compliance.

7. Performance Characteristics