is there a way to use gemini 3.1 live preview with cartesia voices?

Thanks for reaching out, Trevor. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

but does this work with the live model

Thanks — great clarification request. Yes: the Gemini 3.1 Flash Live Preview (the live/flash preview model) can be used with Cartesia voices in LiveKit Agents. • You can configure the AgentSession to use the Gemini 3.1 live model as the LLM and Cartesia (e.g.,

cartesia/sonic-3

) for TTS. • Limitation to note: the Gemini 3.1 Flash Live Preview has restrictions on mid-session control functions (for example,

generate_reply()

,

update_instructions()

,

update_chat_ctx()

and some tool-calling features). Voice I/O and Cartesia TTS operate normally. Would you like a minimal AgentSession code example wired to the Gemini 3.1 live model and Cartesia TTS or help applying this to your project? , with interactive elements

how would you update these args: ctx, *args, **kwargs, llm=google.beta.realtime.RealtimeModel( model="gemini-3.1-flash-live-preview", voice="Sadachbia", temperature=0.7, ),

Thanks — I’m responding to your request about how to update the args to use the Gemini 3.1 Flash Live Preview with the

Sadachbia

voice. Use the same

ctx, *args, **kwargs

signature but pass the Gemini 3.1 realtime model in

llm

and avoid removed/unsupported fields. Example patterns: Python

def handle_turn(ctx, *args, **kwargs):
    session = AgentSession(
        llm=google.beta.realtime.RealtimeModel(
            model="gemini-3.1-flash-live-preview",
            voice="Sadachbia",
            temperature=0.7,
            thinking_config=types.ThinkingConfig(thinking_level="minimal"),
        ),
        # other session config (tts, stt, vad, etc.)
    )
    # use ctx, args, kwargs as needed for history/seeding

TypeScript

function handleTurn(ctx: any, ...args: any[]) {
  const session = new AgentSession({
    llm: new google.beta.realtime.RealtimeModel({
      model: "gemini-3.1-flash-live-preview",
      voice: "Sadachbia",
      temperature: 0.7,
      thinkingConfig: { thinkingLevel: "minimal" },
    }),
    // other session config (tts, stt, vad, etc.)
  });
  // use ctx and args for seeding history or input
}

Key notes: • Remove unsupported params like

enable_affective_dialog

and

proactivity

. • Replace

thinkingBudget

with

thinkingLevel

via

thinking_config

/

thinkingConfig

. •

generate_reply()

,

update_instructions()

, and

update_chat_ctx()

are not compatible mid-session. •

send_client_content

is only allowed for initial history seeding; subsequent uses will be rejected. If you want, tell me which language (Python/TS) and where you currently pass

ctx

,

args

, and

kwargs

and I’ll adapt the snippet to your exact function. Sources: Gemini Live API plugin | LiveKit Documentation, with interactive elements

but I want to use Cartesia voices

Thanks — understood. You can use Gemini 3.1 Flash Live Preview as the LLM while using Cartesia for TTS. Below are minimal examples showing how to pass

ctx, *args, **kwargs

and wire Gemini 3.1 to Cartesia voices. Pick Python or TypeScript and I can tailor it to your code. Python example

def handle_turn(ctx, *args, **kwargs):
    session = AgentSession(
        llm=google.beta.realtime.RealtimeModel(
            model="gemini-3.1-flash-live-preview",
            # Gemini voice setting can be omitted since Cartesia will handle TTS
            temperature=0.7,
            thinking_config=types.ThinkingConfig(thinking_level="minimal"),
        ),
        tts="cartesia/sonic-3",  # Cartesia TTS provider
        # other agent config (stt, vad, turn detection...)
    )

    # seed initial history using ctx/args/kwargs if needed (only allowed at session start)
    session.seed_history(ctx)
    # handle turn input from args/kwargs and run session
    return session.run_turn(*args, **kwargs)

TypeScript example

function handleTurn(ctx: any, ...args: any[]) {
  const session = new AgentSession({
    llm: new google.beta.realtime.RealtimeModel({
      model: "gemini-3.1-flash-live-preview",
      temperature: 0.7,
      thinkingConfig: { thinkingLevel: "minimal" },
    }),
    tts: "cartesia/sonic-3",
    // other session config
  });

  // seed initial history with ctx (allowed only at session start)
  session.seedHistory(ctx);
  return session.runTurn(...args);
}

Key points • Use

tts

set to the Cartesia model (e.g.,

cartesia/sonic-3

or a specific voice ID) in the AgentSession. • Don’t include unsupported Gemini 3.1 fields (

enable_affective_dialog

,

proactivity

). Use `thinking_config`/`thinkingConfig` instead of

thinkingBudget

. •

send_client_content

is allowed only for initial history seeding;

generate_reply()

,

update_instructions()

, and

update_chat_ctx()

are not compatible mid-session. • Cartesia handles TTS output; Gemini 3.1 handles the LLM responses. Want me to: • Insert a specific Cartesia voice ID (tell me the voice ID), or • Adapt the snippet to your existing function (share the function signature)?, with interactive elements

I want to use 1fc31370-81b1-4588-9c1a-f93793c6e01d

Great — thanks for the voice ID. Below are minimal Python and TypeScript examples showing how to wire Gemini 3.1 Flash Live Preview as the LLM and use the Cartesia voice ID

1fc31370-81b1-4588-9c1a-f93793c6e01d

for TTS while keeping your function signature

ctx, *args, **kwargs

. Python

from livekit.plugins import google, cartesia, types

def handle_turn(ctx, *args, **kwargs):
    session = AgentSession(
        llm=google.beta.realtime.RealtimeModel(
            model="gemini-3.1-flash-live-preview",
            temperature=0.7,
            thinking_config=types.ThinkingConfig(thinking_level="minimal"),
        ),
        tts="cartesia/sonic-3",                     # Cartesia model
        tts_voice_id="1fc31370-81b1-4588-9c1a-f93793c6e01d",  # specific voice ID
        # other config: stt, vad, turn detection...
    )

    # seed initial history with ctx (only allowed at session start)
    session.seed_history(ctx)
    return session.run_turn(*args, **kwargs)

TypeScript

import * as google from '@livekit/agents-plugin-google';

function handleTurn(ctx: any, ...args: any[]) {
  const session = new AgentSession({
    llm: new google.beta.realtime.RealtimeModel({
      model: "gemini-3.1-flash-live-preview",
      temperature: 0.7,
      thinkingConfig: { thinkingLevel: "minimal" },
    }),
    tts: "cartesia/sonic-3",                     // Cartesia model
    ttsVoiceId: "1fc31370-81b1-4588-9c1a-f93793c6e01d", // specific voice ID
    // other config: stt, vad, turn detection...
  });

  session.seedHistory(ctx); // allowed only at session start
  return session.runTurn(...args);
}

Notes • Omit unsupported Gemini 3.1 params (

enable_affective_dialog

,

proactivity

). Use `thinking_config`/`thinkingConfig` instead of

thinkingBudget

. •

generate_reply()

,

update_instructions()

, and

update_chat_ctx()

are not compatible mid-session. •

send_client_content

may only be used for initial history seeding. • If your SDK uses different field names for specifying a Cartesia voice, swap

tts_voice_id

/

ttsVoiceId

for the correct property (I used those names as common patterns). Want me to adapt this to your exact function body or your SDK’s field names? , with interactive elements