Hi LiveKit team, I found a working solution for push to talk on the openai realtime and wanted to share our findings in case it's helpful for documentation or API improvements. The Problem(s) 1. Using server_vad with create_response=False produced multiple interim transcripts during recording, even though we only wanted a single final transcript when the user released the PTT button. 2. turn_detection=None on RealtimeModel - Audio was streamed but never processed. OpenAI didn't know what the user said, responses were generic. 3. turn_detection="manual" on AgentSession alone - Without configuring the RealtimeModel, OpenAI's built-in VAD would auto-respond mid-speech. 4. Just commit_user_turn() without generate_reply() - With turn_detection=None, no response was generated at all. The agent stayed in "listening" state. The Solution: We had to bypass the AgentSession's commit_user_turn() and instead directly access the underlying OpenAI Realtime session to send native events: # Configuration session = AgentSession( llm=openai.realtime.RealtimeModel( turn_detection=None, # Disable OpenAI's VAD entirely ), turn_detection="manual", ) # On button release rt_session = session._activity.realtime_llm_session rt_session.commit_audio() # input_audio_buffer.commit rt_session.send_event({"type": "response.create"}) This mirrors how we handle PTT in our direct iOS/WebRTC OpenAI implementation and produces a single clean transcript. Feedback: 2. PTT helper method - Something like session.commit_and_respond() that wraps commit_audio() + response.create would simplify this common pattern. 3. Documentation - The PTT examples focus on STT/LLM/TTS pipelines. A Realtime API-specific example showing this direct commit approach would be valuable. 4. Clarify turn_detection=None behavior - It wasn't obvious that this disables transcription processing entirely, not just auto-response. I'd love to talk if you think there's a better solution or if you want to understand my use case better Thanks for the help!

Thanks for the feedback—I'll check for you & see if there's another recommended practice