Phone agent turn-taking: VAD plus transcript completeness beats silence timers alone
Kindly-Duty272 · reddit · 2026-10-03
The author drilled into one narrow problem: when a phone agent should start talking after the caller stops. Too fast cuts people off; too slow feels dead.
- VAD threshold tuning alone wasn't enough: people pause mid-sentence, and a silence timer can't tell "I'm done" from "let me think."
- What worked: combining VAD with STT partial transcripts — if the text ends on an incomplete construction (e.g. a trailing conjunction), the agent waits even during silence; complete thoughts trigger faster responses.
- TTS latency must be tuned together with end-of-turn logic: a slow voice start reads as a mistake, not a natural beat, regardless of whether you run a Deepgram-style WebSocket pipeline.
Most tutorials treat turn-taking as a pure audio problem; the language signal matters just as much.
More from coding & agent
- Vibe-coding a Japanese town with his daughter: Opus 5.5 outshines Astra — Merzmensch · 2026-10-04
- virattt's earnings KPI pipeline is now 1.8x faster, queryable in ~21s — virattt · 2026-10-04
- Meta paper: RL post-training hurts test-time scalability — the 'Sharpening Tax' — dair_ai · 2026-10-04
- AutoCompact trains agents to compact context themselves, +9.2 on SWE-bench Verified — omarsar0 · 2026-10-04
- Professor says AI grading passed his confidence test: TAs may never grade again — ipeirotis · 2026-10-04
- universal-modder: an open-source kit that lets AI coding agents mod almost any PC game, 2,300 stars in 3 days — alex_verem · 2026-10-04