If you have worked with Asterisk AMD (answering machine detection), you know the default settings are way too aggressive. We were losing about 12% of live calls because short greetings like "hello?" were getting classified as machine beeps. Here is what is actually happening under the hood: most AMD algorithms measure initial silence duration and first voice segment length. When someone picks up and says a short "hello?" followed by quick silence, the algorithm sees a pattern similar to a voicemail beep. Live call gets dumped. Three changes got us from 12% false positives to under 3%: - Extended initial silence threshold from 2500ms to 3500ms - Increased minimum word count before disposition from 2 to 4 - Added a "human detected" override for utterances shorter than 1.5s followed by silence (the classic "hello?" pattern) The full parameter tuning process with specific values is documented at vicistack.com/blog/vicidial-amd-false-positive-reduction/ — tested across about 2 million calls. Key insight: optimizing for false positives matters more than detection speed. A 1-second delay connecting a live call costs way less than dropping it entirely. Anyone else tuning AMD? What parameters are you running?