Server Events

The server sends these JSON events as text frames on the Realtime WebSocket. Each event has a type field. Event names follow the OpenAI Realtime GA protocol.

Session

Type	When	Key fields
`session.created`	Socket is ready	`session`
`session.updated`	A `session.update` was applied	`session`

Input audio

Type	When	Key fields
`input_audio_buffer.speech_started`	Speech detected. Stop local playback here (barge-in)	`audio_start_ms`, `item_id`
`input_audio_buffer.speech_stopped`	Speech ended	`audio_end_ms`, `item_id`
`input_audio_buffer.committed`	The user turn was committed	`item_id`
`input_audio_buffer.cleared`	The input buffer was cleared

Conversation items

Type	When	Key fields
`conversation.item.added`	An item joined the history	`item`
`conversation.item.done`	An item finished	`item`
`conversation.item.input_audio_transcription.delta`	Partial user transcript	`item_id`, `delta`
`conversation.item.input_audio_transcription.completed`	Final user transcript	`item_id`, `transcript`

Responses

Type	When	Key fields
`response.created`	A reply started	`response`
`response.output_item.added`	An output item started	`response_id`, `item`
`response.output_item.done`	An output item finished	`response_id`, `item`
`response.content_part.added`	A content part started	`item_id`, `part`
`response.content_part.done`	A content part finished	`item_id`, `part`
`response.done`	The reply finished. Carries usage	`response` (`status`, `usage`)

Audio output

Type	When	Key fields
`response.output_audio.delta`	A chunk of reply audio	`response_id`, `item_id`, `delta` (base64)
`response.output_audio.done`	Reply audio finished	`response_id`, `item_id`
`response.output_audio_transcript.delta`	A chunk of reply transcript	`response_id`, `item_id`, `delta`
`response.output_audio_transcript.done`	Reply transcript finished	`response_id`, `item_id`, `transcript`
`output_audio_buffer.started`	Output audio playback started
`output_audio_buffer.stopped`	Output audio playback stopped
`output_audio_buffer.cleared`	Output audio buffer cleared, for example after barge-in

Decode each delta from base64 and play the chunks in order at the configured output rate. The default is PCM16 at 24 kHz, mono.

Text output

Type	When	Key fields
`response.output_text.delta`	A chunk of reply text	`response_id`, `item_id`, `delta`
`response.output_text.done`	Reply text finished	`response_id`, `item_id`, `text`

Function calls

Type	When	Key fields
`response.function_call_arguments.delta`	Streaming tool arguments	`call_id`, `delta`
`response.function_call_arguments.done`	Tool arguments complete	`call_id`, `name`, `arguments`

See Function tools for the reply loop.

Errors

Type	When	Key fields
`error`	A request or session failed	`event_id`, `error` (`type`, `code`, `message`, `param`)

Examples

{
  "type": "response.done",
  "response": {
    "status": "completed",
    "usage": {
      "total_tokens": 96,
      "input_tokens": 42,
      "output_tokens": 54
    }
  }
}

{
  "type": "error",
  "event_id": "event_...",
  "error": { "type": "invalid_request_error", "code": "invalid_value", "message": "...", "param": "session.audio.output.voice" }
}

Get Started

Guides

Reference

Server Events

Session

Input audio

Conversation items

Responses

Audio output

Text output

Function calls

Errors

Examples

​Session

​Input audio

​Conversation items

​Responses

​Audio output

​Text output

​Function calls

​Errors

​Examples

​Related

Session

Input audio

Conversation items

Responses

Audio output

Text output

Function calls

Errors

Examples

Related