how download-files work?

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

if my python file is like this, what will it download:

from livekit.agents import AgentSession, Agent, RoomInputOptions, ModelSettings, metrics, MetricsCollectedEvent
from livekit.plugins import (
    openai,
    cartesia,
    elevenlabs,
    deepgram,
    noise_cancellation,
    silero,
)
from livekit.plugins.turn_detector.multilingual import MultilingualModel
from livekit.plugins.turn_detector.english import EnglishModel

Thanks — I’m responding to which model files will be downloaded given the imports in your Python file. Based on those imports, download-files will fetch model files required by the plugins that need local models: • noise_cancellation: downloads the noise-cancellation model files used by the

livekit-plugins-noise-cancellation

plugin. • silero: downloads Silero VAD model weights used by

livekit-agents[silero]

. • turn_detector (MultilingualModel and EnglishModel): downloads the turn-detection model files for both the multilingual and English turn-detector implementations used by

livekit-agents[turn-detector]

. The other imported plugins do not require local model weights to be downloaded by `download-files`: • openai, cartesia, elevenlabs, deepgram — these rely on external APIs/SDKs and do not add local ML model downloads. If you want, tell me whether you expect multilingual+English turn models or only one, and I’ll list the exact model filenames and a sample command to download them.

I want both multilingual+English turn models, give me a python code

You're asking for a Python file that will cause the download of both the multilingual and English turn-detection models when you run the download step. Below is a concise example that imports and references both models so

uv run agent.py download-files

will fetch them (plus Silero VAD and noise-cancellation if those plugins are present). Save as agent.py:

# agent.py
from livekit.agents import AgentSession
from livekit.plugins.turn_detector import MultilingualModel, EnglishModel
from livekit.plugins import silero, noise_cancellation

# Reference both models so download-files will include them
multilingual_model = MultilingualModel()
english_model = EnglishModel()

# Optional: configure an AgentSession that uses one of the models at runtime
session = AgentSession(
    turn_detection=multilingual_model,  # use multilingual at runtime (recommended)
    vad=silero.VAD.load(),              # ensures Silero VAD model is referenced
    # add other components (stt, tts, llm) as needed
)

if __name__ == "__main__":
    print("Agent module loaded. Run `uv run agent.py download-files` to download models.")

Then run:

uv run agent.py download-files

This will download the multilingual and English turn-detector models as well as any models required by

silero

and

noise_cancellation

plugins referenced in the file. Sources: Improved End-of-Turn Model Cuts Voice AI Interruptions 39% | Building the all-in-one platform for voice AI agents | LiveKit turn detector plugin | LiveKit docs | Voice AI quickstart | LiveKit docs

do I have to initialize like this:

multilingual_model = MultilingualModel()
english_model = EnglishModel()

Good question — short answer: no, you don't strictly have to initialize both with

MultilingualModel()

and

EnglishModel()

just to trigger downloads. Details: •

uv run agent.py download-files

downloads model files for plugins you import and reference. Explicitly instantiating a model (e.g.,

MultilingualModel()

) ensures that model is referenced and its weights are included. • The multilingual model largely replaces the English model and is recommended. If you only need multilingual at runtime, instantiating just

MultilingualModel()

is sufficient for runtime use;

download-files

may still fetch both models if the turn-detector plugin registers them as available. • Best practice: explicitly reference any model you definitely want downloaded. If you want both multilingual and English files present, instantiate both as you showed. If you only need multilingual, instantiate only that one.

are you sure I just have to import?

Short answer: importing alone is not guaranteed to download the model files. Explanation and recommendation: • The

download-files

step downloads files for plugins and models that are actually referenced by your code. Simply importing a module may not register the specific model class for download. • To be certain a model's weights are fetched, explicitly reference or instantiate it (e.g.,

MultilingualModel()

and/or

EnglishModel()

), or otherwise use the class in your code so the download step can detect it. • If you want both multilingual and English turn models present, instantiate both before running:

multilingual_model = MultilingualModel()
english_model = EnglishModel()

If you tell me whether you prefer only multilingual at runtime (but still want both downloaded) or you want both active, I’ll give the minimal code snippet for that case.