Speech and turn detection

config.stt is one {provider, model} pair from the catalog’s stt array, plus an optional fallback. It decides how the caller is transcribed and what ends their turn: there is no turn-detection setting, it follows from the model and config.language. Voice and video cascade agents; a text agent’s optional stt only transcribes voice notes. There is no language on stt.

{ "language": "en", "stt": { "provider": "openai", "model": "gpt-transcribe" } }

Streaming or batch

The catalog entry’s streaming flag decides it.

	Streaming (`streaming: true`)	Batch (`streaming: false`)
How it works	A socket transcribes as the caller speaks	A local voice detector cuts utterances; each is one request
Reply latency	Transcript ready when the turn ends	Adds a request round-trip
Billing	The open socket, for the whole call	Only the utterances sent
Interim transcripts	Yes	No
`interruption.min_words`	Works	No effect (publish warns)
What ends the turn	The model’s own end-of-speech detector	The end-of-turn model, or silence

A fallback must have the same streaming value as the primary. OpenRouter speech-to-text models are all batch.

What ends a turn

Pipeline	Detector
Streaming model	The provider’s own end-of-speech detector.
Batch model + `language` in `turn_detector_languages`	LiveKit’s local end-of-turn model, which judges whether the sentence sounds finished and waits through a mid-thought pause. No extra cost.
Batch model + any other language, or Auto	Silence alone: the turn ends after `min_silence_duration`.

GET /v1/catalog returns turn_detector_languages (14 languages, including English, Hindi, Spanish, Chinese and Arabic). Matching uses the primary subtag, so en-IN and pt-BR qualify. A batch model with one of those languages is the only pipeline that judges meaning rather than silence. A realtime model detects turns inside the provider; see realtime models.

Endpointing per model

turn_handling.endpointing has two fields (defaults and limits in turn handling):

max_silence_duration is used only by the end-of-turn model, so only on a batch model with a covered language. Elsewhere it is ignored.
min_silence_duration always applies, but what it controls depends on the model:

Model	What `min_silence_duration` is
Any batch model	The exact silence wait, via the local voice detector.
`deepgram/nova-3*`, `xai/xai-stt`	Sent as the provider’s end-of-speech window.
`openai` streaming	Sent to the voice detector that commits audio to OpenAI.
`elevenlabs` streaming	Sent to the provider, clamped to 0.3–3.0 s.
`soniox`	Sent to the provider, clamped to 0.5–3.0 s.
`deepgram/flux-`, `sarvam/`	A floor only; these endpoint on their own schedule.

The same value tunes the local voice detector for the whole call.

Fixing mistimed turns

Symptom	Fix
Cuts people off mid-sentence	Move to a batch model with `language` set to a covered language; on streaming, raise `min_silence_duration`.
Still cuts off on a batch model	`language` is Auto or not covered. Set a covered code.
Cuts off people who trail off, sometimes	Raise `max_silence_duration`.
Long dead air before each reply	Lower `min_silence_duration`, or use a streaming model.
Raising `min_silence_duration` changes nothing	The model treats it as a floor (Flux, Saaras). Change model.
Ignores a caller talking over it	`min_words` on a batch model. Use `min_speech_duration`. See turn handling.
Transcribed in the wrong language or script	The model does not detect on Auto. See language.

Language

How Auto behaves per model.

Turn handling

Interruptions, endpointing defaults and preemptive generation.

Get started

Build

Guides

Speech and turn detection

Streaming or batch

What ends a turn

Endpointing per model

Fixing mistimed turns

Next

Language

Turn handling

​Streaming or batch

​What ends a turn

​Endpointing per model

​Fixing mistimed turns

​Next

Language

Turn handling

Streaming or batch

What ends a turn

Endpointing per model

Fixing mistimed turns

Next