Connection keepalive

Overview

In real-time transcription, you may have periods where you are not sending any audio frames — for example when using client-side VAD (voice activity detection), during pauses in speech, or when you intentionally stop streaming audio.

To keep the session alive and preserve context, you must send a keepalive control message:

{"type": "keepalive"}

This prevents the WebSocket connection from timing out when no audio is being sent.

When to use

Send a keepalive message whenever:

• You only stream audio during speech (client-side VAD).
• You temporarily pause audio streaming but want to keep the session active.

This ensures that:

• The connection stays open.
• Session context (e.g., speaker labels, language tracking, prompt) is preserved.

Key points

• Send at least once every 20 seconds when not sending audio.
• You may send more frequently (every 5–10s is common).
• If no keepalive or audio is received for 20 seconds, the stream ends with a 408 request_timeout error, the connection is closed, and the stream stops counting against your concurrent requests limit.
• Before the start request, send a keepalive at least every 40 seconds. See below.

Before the start request

A connection can be opened early and held: send a keepalive at least every 40 seconds until the start request, or the connection is closed as idle with WebSocket status 1001. There is no deadline for the start request itself. A held connection counts against your connection limit.

See WebSocket authentication. A connection that still sends api_key in the start request follows the rules in Migrate WebSocket authentication.

• You are charged for the full stream duration, not just the audio processed.