Tap to start a voice chat.
Same knobs as the old OpenAI Realtime page: server VAD + barge-in. Grok has no far_field NR field — we enable Grok noise suppression + browser AEC instead. Reasoning is off for low latency.
Speakerphone: higher VAD + barge-in. Headphones: lower threshold. Both keep the mic live so you can interrupt.
Server VAD ends your turn after silence. Off = manual (no auto response).
Higher = needs louder speech (0.1–0.9). OpenAI speakerphone default was 0.72.
Audio included before speech start — stops clipping the first syllable.
OpenAI browser default was 550ms. Higher = more thinking pauses before the agent replies.
Maps to Grok turn_detection.interrupt_response. Local playback is also cleared on speech start.
xAI enable_noise_suppression (closest thing to OpenAI far_field on this API).