Source: OpenAI Realtime Voice Activity Detection Guide

OpenAI's guide to turn detection in Realtime sessions: server_vad parameters, semantic_vad with eagerness, events and trade-offs.

Contents
  • What: official API documentation (Voice activity detection)
  • Author/publisher: OpenAI
  • Date: living document; read 2026-11-14
  • Original: Voice activity detection

Summary

Explains the two automatic turn-detection modes of the Realtime API and how to tune them.

Key claims

  • VAD is on by default in speech-to-speech sessions; configured at session.audio.input.turn_detection; default mode server_vad.
  • server_vad is silence-based: threshold (0-1, higher for noisy rooms), prefix_padding_ms, silence_duration_ms (shorter = faster turns, more cut-offs), plus create_response and interrupt_response.
  • semantic_vad uses a classifier on the user’s words to score whether they have finished; low probability waits for a timeout. eagerness = low / medium (auto, the default) / high tunes how long it waits (the guide gives no numeric wait times).
  • Semantic VAD makes premature interruptions and split transcripts less likely.
  • Events: input_audio_buffer.speech_started and input_audio_buffer.speech_stopped.
  • Some transcription models require turn detection off (null), with the client sending input_audio_buffer.commit manually.

Feeds

Voice agents

Suggest changes →