Skip to main content
The Stream API provides real-time speech-to-speech translation over a single persistent WebSocket connection. It supports both consecutive mode (speak, wait, hear translation) and simultaneous mode (continuous translation as you speak). This page covers the WebSocket transport: endpoint, authentication, audio frames, connection errors, and connection lifetime. For the shared application message layer, including init, configure, flush, and server events, see the Stream message layer.

​
Endpoint

​
Authentication

Authentication is provided via query parameters. Provide one of the following: Example connection URLs:
Authentication is validated before the WebSocket upgrade. On failure the server returns HTTP 401, not a WebSocket close frame.

​
Transport Model

  • Send Stream JSON messages as WebSocket text frames.
  • Send input audio as raw PCM bytes in WebSocket binary frames after receiving configured.
  • Receive output audio as raw PCM bytes in WebSocket binary frames.
  • Do not wrap audio in JSON or base64.
For message schemas, usage modes, and examples, see the Stream message layer.

​
Connection Lifecycle

​
Audio Frames

Input audio and output audio use raw PCM binary frames on the WebSocket transport.
  • Encoding: 16-bit signed integer PCM (little-endian)
  • Channels: Mono (1 channel)
  • Input sample rate: Must match input_sample_rate from init
  • Output sample rate: Matches output_sample_rate from init
Send audio in small frames, typically 10-20 ms of audio per frame, for the lowest latency. Larger frames add latency equal to the frame duration.

​
Error Handling

​
Connection Errors

If authentication fails, the server returns an HTTP error before the WebSocket upgrade:

​
Runtime Errors

During a session, runtime errors are delivered as error messages. Non-fatal errors, such as bad JSON or an unknown message type, do not close the connection. Fatal backend errors may be followed by the connection closing. See error for the message shape.

​
Inactivity Timeout

If no messages are sent or received for 60 seconds (configurable), the server closes the WebSocket connection.

​
Migration

For the v2 consecutive API to v3 Stream API migration guide, see Migration Guide: v2 Consecutive API to v3 Stream API.