The Eleven Realtime provider profile — the knobs the shared OpenAIRealtime protocol
implementation runs with. Each gate is read off ElevenLabs' published session schema
(OpenAiSessionConfigInput / OpenAiSessionConfigOutput in @elevenlabs/elevenlabs-js 2.74.0):
No input-transcription model. The models are cascaded, so the user's speech is transcribed
by their own ASR stage; audio.input.transcription takes only a language, and sending
OpenAI's model key there would be an unknown field.
No reasoning effort, no parallel tool calls. Both are accepted on input but absent from the
session the server echoes back, i.e. accepted for SDK compatibility and ignored. Max reasons
internally; how hard is not a session knob. Sending them would be the transmitted-then-discarded
pattern #3374 exists to prevent.
MCP tools and an output voice: yes. Both survive into the echoed session. The voice is an
ElevenLabs voice id.
server_vad and semantic_vad, with create_response, so meeting mode and live
reconfiguration both have a field to drive.
The Eleven Realtime provider profile — the knobs the shared OpenAIRealtime protocol implementation runs with. Each gate is read off ElevenLabs' published session schema (
OpenAiSessionConfigInput/OpenAiSessionConfigOutputin@elevenlabs/elevenlabs-js2.74.0):audio.input.transcriptiontakes only alanguage, and sending OpenAI'smodelkey there would be an unknown field.server_vadandsemantic_vad, withcreate_response, so meeting mode and live reconfiguration both have a field to drive.