Client Events - 60db

Send these JSON events as text frames on the Realtime WebSocket. Field names and shapes follow the OpenAI Realtime GA protocol. Each event has a type field and may include an optional event_id.

Session

Type	Purpose	Fields
`session.update`	Change the session config. Send before streaming audio	`session` (object): `type` (`realtime`), `model`, `output_modalities`, `instructions`, `audio`, `tools`, `tool_choice`, `memory`, `knowledge`, `intent`

{
  "type": "session.update",
  "session": {
    "type": "realtime",
    "output_modalities": ["audio"],
    "instructions": "You are a friendly voice assistant.",
    "audio": {
      "input": { "turn_detection": { "type": "server_vad", "silence_duration_ms": 1000 } },
      "output": { "voice": "YOUR_VOICE_ID" }
    }
  }
}

Input audio buffer

Type	Purpose	Fields
`input_audio_buffer.append`	Add audio to the input buffer	`audio` (string): base64 audio in the configured input format
`input_audio_buffer.commit`	End the current user turn. Needed in manual turn mode	none
`input_audio_buffer.clear`	Discard buffered audio that is not committed	none

{ "type": "input_audio_buffer.append", "audio": "<base64 PCM16 mono, 24 kHz>" }

Conversation items

Type	Purpose	Fields
`conversation.item.create`	Add a message or a function result to the history	`item` (object), optional `previous_item_id`
`conversation.item.retrieve`	Fetch an item	`item_id`
`conversation.item.truncate`	Cut off assistant audio the user did not hear	`item_id`, `content_index`, `audio_end_ms`
`conversation.item.delete`	Remove an item from the history	`item_id`

Add a user text message:

{
  "type": "conversation.item.create",
  "item": {
    "type": "message",
    "role": "user",
    "content": [{ "type": "input_text", "text": "I prefer short answers." }]
  }
}

Return a function result:

{
  "type": "conversation.item.create",
  "item": { "type": "function_call_output", "call_id": "call_abc123", "output": "{\"status\":\"shipped\"}" }
}

Responses

Type	Purpose	Fields
`response.create`	Start a model reply now	optional `response` (object) with per-response overrides
`response.cancel`	Stop a reply in progress	optional `response_id`

Out-of-band responses (response.create with conversation: "none") are not supported yet.

Ordering

Send session.update before the first input_audio_buffer.append.
Send response.create after you add a function result. With server_vad, the server also starts replies on its own when the user stops speaking.
In manual turn mode, send input_audio_buffer.commit and then response.create for each turn.