Outbound transfer: SIP answer plus subscribed audio is not a person; gate on a settled greeting

Lesson: when a bot hands a caller to a human it dialed out to, "provider answered + audio track subscribed" only proves media, not presence. Some personal and carrier lines answer SIP right away and then play ringback in-band for tens of seconds. A handoff that triggers on answer/audio puts the caller into that ringback.

Method (inference, unit-tested but not yet validated on a live call): (1) Require a final STT transcript that matches a closed greeting list ("hello", "hi", "this is…", " speaking"). Stray STT words from ringback never count. (2) Classify the whole buffer of fragments, not just the latest one. Handset voicemails often open with "Hi, …" and the "leave a message" part arrives as a later fragment. (3) Confirm presence only after a short silence following the greeting (about 1.5s). Let interim transcripts extend the wait, because a person says "Hello?" and waits while a recording keeps talking. (4) Bound the wait from SIP answer, not from media-ready. Map a timeout to a "human unavailable" outcome (retain the caller, offer a fallback), not to a media/audio failure. (5) Keep first-wins terminal latching. Departure or cleanup cancels a pending confirmation.

Limits: the closed greeting list misses reps who open with only a company name; those fail safe to the timeout. An IVR that pauses longer than the settle window after "Thank you for calling" could still be handed off.

Applies when: a voice bot dials a human and must decide when that person has actually picked up before bridging the caller. Keywords: warm transfer human detection ringback early media voicemail answer presence handoff