Start real-time speech-to-text on a leg

Language code (e.g. "en", "es")

Emit partial (non-final) transcripts

STT provider: "elevenlabs" (default), "deepgram" (/v1/listen), "deepgram_flux" (/v2/listen, conversational turn detection), "azure" or "speechmatics"

API key override (falls back to ELEVENLABS_API_KEY, DEEPGRAM_API_KEY, AZURE_SPEECH_KEY or SPEECHMATICS_API_KEY env var depending on provider)

Provider-specific model. Deepgram: default "nova-3". Deepgram Flux: "flux-general-en" (default) or "flux-general-multi". Speechmatics: "standard" (default) or "enhanced".

Deepgram: milliseconds of silence before a segment is finalized; 0 disables endpointing. Speechmatics: maps to max_delay, clamped to 700-4000 ms; 0 leaves the provider default.

Deepgram: milliseconds of silence after which an stt.turn event with event=utterance_end is emitted. Deepgram requires interim results for this, which are requested automatically and still suppressed unless partial is true. Speechmatics: milliseconds of silence that close a turn and emit an stt.turn event with event=end_of_turn — default 600, capped at 2000, and 0 disables turn detection.

Deepgram Flux only: end-of-turn confidence that fires an eager_end_of_turn stt.turn event, enabling speculative generation. Must be between 0.3 and 0.9. When unset, no eager_end_of_turn or turn_resumed events are emitted at all.

Deepgram Flux only: end-of-turn confidence required to close a turn. Deepgram default 0.7.

Deepgram Flux only: milliseconds of silence after which a turn is closed regardless of confidence. Deepgram default 5000.

Instance identifier