Tap to start a voice chat.
Industry fix for speaker + barge-in: browser AEC + OpenAI far_field noise reduction + tuned VAD — not muting your mic.
Speakerphone: far-field NR + higher VAD + barge-in on. Headphones: near-field + sensitive VAD. Both keep the mic live so you can interrupt.
far_field is what mobile devs use for speakerphone on Realtime API — filters audio before VAD so echo is less likely to count as “you speaking.”
WebRTC subtracts speaker audio from the mic using the playback reference. system uses OS AEC when available (often better on iPhone).
Server VAD ends your turn after silence. “Off” keeps the mic open until you mute or end the call (you may need to pause longer for the model to reply).
Higher = needs louder speech to start a turn. Helps in noisy rooms; too high may miss soft speech.
Audio included before detected speech. Stops clipping the first syllable.
How long you must be quiet before the model treats your turn as finished. Lower = snappier; higher = better for thoughtful pauses.
Only if echo still wins after far_field + AEC. Blocks real barge-in while the agent talks — use speakerphone preset first.
How ChatGPT voice works: your speech can cut the agent off once echo is under control (AEC + far_field).
OpenAI realtime output voice. Takes effect on the next new call.