Conversation tuning
Choose the model, the voice speed and how eagerly the bot takes its turn
Every knob below is optional and per bot; leave it out and the server default applies.
The model
{
"meeting_url": "...",
"prompt": "...",
"llm_provider": "anthropic",
"llm_model": "claude-opus-4-7"
}| Field | Values | Default |
|---|---|---|
llm_provider | openai, anthropic, zai | the server's LLM_PROVIDER (Anthropic on the hosted API) |
llm_model | any model name the provider accepts | the server's model for that provider (claude-haiku-4-5 on the hosted API) |
The provider must have a key: the server's own, or one you pass in credentials. Asking for a provider the server has no key for returns a 400 that names the missing one.
The defaults are the fast models on purpose: a spoken reply of a sentence or two gains little from a frontier model and waits seconds longer for it. Reasoning is turned off in whatever way the chosen model accepts, so picking a bigger llm_model costs latency but never a failed turn; on models whose reasoning is always on (Claude Fable, Opus 5.5 and newer) the bot runs at the lowest effort instead.
Speech
Text-to-speech is Cartesia; the voice comes from the persona (cartesia_voice_id in its metadata) or the server default. speech_speed is a multiplier between 0.6 and 1.5:
{ "speech_speed": 1.15 }A prompt bot that should sound brisk benefits from both a higher speech_speed and an instruction like "answer in two sentences" — the LLM decides how much is said, the speed only how fast.
Turn-taking
The bot decides that someone has finished talking with voice-activity detection. turn_config tunes it:
{
"turn_config": {
"confidence": 0.5,
"start_secs": 0.25,
"stop_secs": 0.8,
"min_volume": 0.6
}
}| Field | What it does | Default |
|---|---|---|
confidence | How sure the detector must be that it hears speech (0–1) | 0.5 |
start_secs | Sustained speech before a turn registers | 0.25 |
stop_secs | Silence before the bot considers the speaker done and replies | 0.8 |
min_volume | Minimum input level counted as speech (0–1) | 0.6 |
Two patterns cover most needs:
- Humans talking to a bot: keep the defaults, or lower
stop_secsto0.5for snappier replies. - Bots talking to bots: raise
stop_secsto1.2–1.5andstart_secsto0.4so they stop barging in on each other, and read Multiple bots.
Re-engagement when the room is quiet
Once admitted, a bot that hears no new conversation for 15 seconds says something on its own, at most three times, then waits. This is what lets two reactive bots start a conversation without a human, and it stops a bot alone in a room from monologuing. On a self-hosted server the timing is BOT_IDLE_REENGAGE_SECS and the cap BOT_MAX_REENGAGE.
Built-in tools
enable_tools (default true) gives the LLM two small functions, current weather and current time, plus the ability to save a short call summary. Set it to false for a pure conversation; it has no effect on MCP tools, which are configured separately.