Pure guard that tracks sustained speech to trigger barge‑in, exposing a simple API.
// Source: https://github.com/vellum-ai/vellum-assistant | License: MIT | Copyright 2025 Vellum AI // Provided via justclaw.it.com; keep this notice with the code. /** * The sustained-speech barge-in guard, shared by the live-voice session and * the phone media-stream path. * * A single above-gate audio chunk is not an interruption: brief TTS bleed * through imperfect echo cancellation, a cough, a line click. Speech has to * accumulate to `minSpeechMs` before the assistant is cut off. Short gaps * inside a run are tolerated (a browser's half-duplex echo canceller ducks * the near-end voice into intermittent above-gate chunks; a phone line * drops frames), but a longer continuous silence, or a run that is mostly * silence, resets it, so sparse periodic blips can never sum into a ... (truncated -- full source via MCP)
See the full source, get the GitHub permalink, and search for more like it.
Get a free API key