Stream termination

Overview

A real-time stream ends with a well-defined three-step handshake: the client signals the end of input with text_end: true, the server emits a final audio chunk with audio_end: true, and then the server confirms cleanup with terminated: true. Only after terminated: true is it safe to release stream state or reuse the stream_id.

Normal termination sequence

1. Client sends final text chunk with text_end: true:

{
  "text": "Final chunk of text.",
  "text_end": true,
  "stream_id": "stream-001"
}

2. Server sends the last audio chunk with audio_end: true:

{
  "audio": "<base64-audio-bytes>",
  "audio_end": true,
  "stream_id": "stream-001"
}

3. Server sends the final stream event:

{
  "terminated": true,
  "stream_id": "stream-001"
}

• audio_end: true indicates the audio payload sequence is complete, no more audio chunks will arrive for this stream.
• terminated: true indicates stream lifecycle cleanup is complete on the server, it is now safe to release stream state and reuse the stream_id.

Client-initiated cancellation

If you need to stop a stream early, for example, user interruption in a voice agent or a timeout on the calling side — send a cancel message. The server finalizes the stream without sending further audio chunks.

Cancel request:

{
  "stream_id": "stream-001",
  "cancel": true
}

Finalization response:

{
  "terminated": true,
  "stream_id": "stream-001"
}

Error termination

If a stream fails, the server responds with an error message followed by terminated: true for that stream_id:

{
  "stream_id": "stream-001",
  "error_code": 400,
  "error_message": "Missing required field: model"
}

{
  "terminated": true,
  "stream_id": "stream-001"
}

The WebSocket connection stays open unless there is a connection-level failure, so other streams continue. For the full list of error codes and messages, see the WebSocket API reference.