Skip to main content

Voice Activity Detection (VAD)

The Voice Settings group on the configuration screen controls how the agent detects user speech, handles interruptions, and announces tool usage.

VAD parameters (Sync / Async modes)​

VAD detects when the user has started or stopped speaking. The following parameters are editable directly from the voice agent Edit tab:

  • Silence duration (ms) — how long of a silence indicates the user has finished speaking.
  • Model selection — the VAD model used to make the speech-vs-silence decision.
  • Turn end detection type — choose between silence-based and character-count detection. Silence-based is the default; character-count can be useful in noisy environments where silence is unreliable.

VAD parameters (Realtime API mode)​

When the voice agent is in Realtime API mode, the VAD panel shows the Realtime-API-specific turn-detection parameters instead:

  • Threshold — slider, 0.00 – 1.00.
  • Prefix padding — slider, in ms.
  • Silence duration — slider, in ms.
  • Idle timeout — toggle on/off.

Only the Normal detection mode is exposed in Realtime API mode. Semantic turn detection and the ability to disable turn detection entirely are not surfaced today.

Interruption / barge-in​

The Interruption toggle lets users interrupt the agent mid-speech (barge-in). When enabled, the agent stops TTS playback as soon as the user starts speaking and immediately processes the new input.

Combined with VAD, this is what makes a voice agent feel responsive. The settings to tune are:

  • Interruption toggle — on/off.
  • Interruption duration — how quickly the agent reacts once the user starts talking.

Set silence duration and interruption duration together; aggressive interruption with long silence duration can produce false barge-ins.

Tool announcements​

The Tool usage announcements toggle controls whether the agent speaks a short acknowledgement before calling a tool.

  • On: the agent announces tool calls (e.g., "Let me check that for you…") before invoking the tool. Useful when tools have noticeable latency — the user knows the agent is working, not stuck.
  • Off: the agent goes silent during the tool call. Best for fast tools where an announcement adds awkward delay.

The announcement text itself is part of the base instruction / system prompt for the agent.

  • For a customer support voice agent on a phone line: interruption on, silence duration ~700ms, tool announcements on.
  • For a kiosk / mobile voice agent with short queries: interruption on, silence duration ~400ms, tool announcements off.
  • For a Realtime API agent: keep the OpenAI defaults and tune Threshold / Silence duration based on call recordings.

Always validate VAD changes by replaying real call audio in Run history — the audio-synced waterfall makes it easy to see where the agent was incorrectly detecting the end of a turn.