Skip to main content
POST
SetVoiceAgentRealtimeModelAPI chooses how a voice agent turns speech into a

Authorizations

Authorization
string
header
required

An Erdo API key (erdo_api_...) or scoped token (erdo_token_...).

Path Parameters

slug
string
required

Body

application/json
realtime_model
string
required

Response

Success response

agent
string
name
string
realtime_model
string

"mini" and "max" hand the whole turn to a single speech-to-speech model the provider tunes end to end — recognition, turn-taking, the reply and the voice — while the agent's prompt, voices, languages, tools, knowledge and turn settings still apply. Max reasons before it answers and is the one meant for tool-heavy work (booking, knowledge lookups); Mini is cheaper and faster on plain conversation. "off" is the cascade Erdo ran before: the provider's speech recognition, then defaultVoiceAgentLLM, then defaultVoiceAgentTTSModel.

Max is the default (defaultVoiceRealtimeModel), chosen 2026-10-25 after the A/B in docs/voice.md: silence-to-first-audio p50 0.53s against the cascade's 1.78s fleet-wide, at about the same per-token price. "off" is an explicit, stored opt-out.

The values are Erdo's; realtimeModelProviderID is the one place they become provider ids, so nothing outside this file names an ElevenLabs model.

widget
string