AMD listens to the first seconds of a connected call and decides whether a live human answered or a voicemail picked up. Accurate AMD means your agents only hear live humans.
Four layers
- Layer 0: Acoustic fingerprint match (MFCC), known voicemail greetings flagged instantly
- Layer 1: Native Asterisk AMD, silence-based baseline
- Layer 2: Real-time AGI classifier, combines layers 0 and 1 with campaign-specific thresholds
- Layer 3: Post-call AI verification, Whisper transcribes the first seconds, LLM verifies the call
Tuning for your campaign
Admin → AMD → [campaign] lets you set the confidence thresholds. Default values work for most B2C dialing. If you see high false-positive rates (live humans getting voicemail-dropped), lower the confidence. If you see voicemails reaching agents, raise the confidence or enable Layer 3 verification.
Self-learning
When an agent dispositions a call as AM (voicemail), the acoustic fingerprint is saved to the campaign's database. Over a few weeks of dialing, your per-campaign fingerprint database gets very accurate for that audience and carrier mix.