Why “Hello?” breaks default AMD

Picture the most common human answer on an outbound call. The phone rings, there’s the usual dialer pause, and the person jumps in fast: “Hello? Hello, who is this?” No hesitation, no gap. That’s a live human being completely normal — and it’s exactly the answer default AMD is worst at.

The “trailing pause” that default AMD depends on

Stock Asterisk AMD is timing-based. The pattern it’s tuned for: a person answers, says a short greeting, and then stops — waiting for the other side to respond. That trailing pause is the tell. A machine, by contrast, keeps going (“...you’ve reached Bob, please leave a message after the tone”) — a longer, continuous run of speech.

So the rule becomes, roughly: short speech then silence = human; long continuous speech = machine. It works right up until a real person doesn’t pause.

Fast answerers look exactly like machines

Someone who answers rapidly — “Hello? Hello, can you hear me? Who’s this?” without stopping — produces a continuous run of speech with no trailing pause. To a timing-based detector, that’s the signature of a machine greeting. It fires a false positive, classifies the call as voicemail, and hangs up. The person did nothing wrong; they just answered like a normal, eager human. The faster and friendlier they are, the more likely they get dropped.

The TCPA timing bind makes it worse

There’s a second squeeze. Greeting-dependent detection has to wait for the speech to finish before it can decide — but the TCPA / Telemarketing Sales Rule two-second compliance clock starts the moment the consumer completes their greeting. So the fastest answers give the detector the least time to classify, while simultaneously tightening the window you have to connect a live agent. The architecture is fighting itself: it needs more speech to be sure, and the clock is punishing it for waiting.

The fix: decide on the sound, at the moment of answer

The way out is to stop depending on the shape of the greeting at all. A live call has an acoustic signature that arrives at or near the moment of answer — the line-noise profile, the room and background of a real person’s environment, the texture of a live connection — often before anyone has said a full word. Detect on that, and the fast-greeting problem simply dissolves: there’s no pause to wait for, no word count to get fooled by, and no race against the speech finishing.

That’s how AMDY works. It classifies the sound of what answered, not the structure of what was said, so a person who blurts “Hello? Who’s this?” is identified as a human in a fraction of a second — and connected, not hung up on.