Swapping providers

Every provider choice lives in one place: the AgentSession built in agent/src/agent.py.

agent/src/agent.py

session: AgentSession = AgentSession(
    stt=deepgram.STT(model="nova-3"),
    llm=cerebras.LLM(model="gemma-4-31b"),
    tts=inworld.TTS(model="inworld-tts-2", voice="Ashley"),
    vad=ctx.proc.userdata["vad"],
    turn_handling=TurnHandlingOptions(
        turn_detection=MultilingualModel(),
        interruption={"mode": "vad"},
    ),
)

The plugins come from one import at the top of the same file:

agent/src/agent.py

from livekit.plugins import cerebras, deepgram, inworld, silero
from livekit.plugins.turn_detector.multilingual import MultilingualModel

Changing one

Three edits, all in agent/.

Add the plugin

cd agent
uv add livekit-plugins-openai

Import it

Add it to the from livekit.plugins import ... line in src/agent.py.

Replace the argument

Swap the one constructor argument. The other two are untouched.

llm=openai.LLM(model="gpt-4o-mini"),

Set the new provider’s key in the root .env (and in agent/.env if you run the worker by hand). Then rebuild the worker image:

docker compose up -d --build agent

Voice, model and language

Model names are strings on the constructor, so changing a voice or a model version is a one-word edit. inworld.TTS also takes voice, which is "Ashley" by default.

VAD and turn detection are separate

vad and turn_handling are not provider choices in the same sense:

The Silero VAD is loaded once per process in prewarm() and shared across sessions through ctx.proc.userdata["vad"].
Turn detection is LiveKit’s MultilingualModel, and interruption is set to VAD mode.

Changing speech to text or the model does not touch either.

scripts/doctor.py probes Deepgram, Cerebras and Inworld by name. Its provider list is hardcoded, so after a swap it checks a key the session no longer uses. Its agent-name, LiveKit and stack checks stay correct.

Noise cancellation

An opt-in, noted in a comment in src/agent.py:

cd agent
uv add livekit-plugins-noise-cancellation

Then import RoomInputOptions from livekit.agents and noise_cancellation from livekit.plugins, and pass it to session.start:

room_input_options=RoomInputOptions(
    noise_cancellation=noise_cancellation.BVC(),
)

Get started

Customize

Ship it

Swapping providers

Changing one

Voice, model and language

VAD and turn detection are separate

Noise cancellation

​Changing one

​Voice, model and language

​VAD and turn detection are separate

​Noise cancellation

Changing one

Voice, model and language

VAD and turn detection are separate

Noise cancellation