Set Smart Turn timeout on streaming STT

When Smart Turn is on, the model controls when speech_final fires. Set smart_turn_timeout (1–5000 ms) so extended silence still forces utterance-final even if the model keeps scoring "not done" (user walks away, long think pause). Base Smart Turn setup: Enable Smart Turn on streaming speech-to-text. Base connect loop: Stream speech to text over WebSocket.

Add both query params

smart_turn_timeout only applies when smart_turn is set.

wss://api.x.ai/v1/stt?sample_rate=16000&encoding=pcm&interim_results=true&smart_turn=0.7&smart_turn_timeout=3000

| Param | Range | Role | |---|---| | smart_turn | 0.0–1.0 | End-of-turn confidence threshold (enables Smart Turn). | | smart_turn_timeout | 1–5000 ms | Max silence before forcing speech_final. |

Omit smart_turn_timeout for unlimited model control: speech_final fires only when confidence exceeds the threshold. Docs example uses 3000 ms with smart_turn=0.7.

Utterance-final (is_final=true and speech_final=true) fires when confidence clears the threshold, or when the timeout is exceeded. Below-threshold silence stays chunk final (is_final=true, speech_final=false).

Proxy the WebSocket through your backend so the API key stays off the client.

Pitfalls

• Batch REST has no smart_turn or smart_turn_timeout. Streaming WebSocket only.
• Timeout without smart_turn does nothing.
• Without a timeout, a session can hang if the model keeps predicting "not done."
• endpointing is a separate silence knob for non–Smart Turn (and remains available): Set streaming STT endpointing silence.
• Console API credits are separate from SuperGrok's weekly pool on grok.com.