LiveKit Turn Detection Plugin (Vietnamese)

A specialized turn detection plugin for LiveKit Agents, optimized for Vietnamese language contexts. This plugin uses a local ONNX model to predict the end of a user's turn with high accuracy and low latency.

Features

• Vietnamese Optimized: Fine-tuned for Vietnamese conversational patterns.
• Local Inference: Runs ONNX model locally for privacy and speed.
• Context Aware: Considers preceding assistant messages for better prediction accuracy.
• Lightweight: Uses SafeTensor / ONNX quantization for minimal memory footprint.

Development Workflow

flowchart TD
    Train[Train BERT Model] -->|Export| Export[Export to ONNX]
    Export -->|Copy & Replace| Replace[Replace livekit/plugins/turn_detection/model_quantized.onnx]
    Replace -->|Build| Build[Build Wheel .whl]
    Build -->|Push| Push[Push to S3 or PyPI]
    Push -->|Install via uv/pip| Install[Install in Target Project]

Build Instructions

You can build the package wheel using uv (recommended) or build.

Build Wheel

Run the build command in the project root:

uv build

This will create a .whl file in the dist/ directory (e.g., dist/livekit_plugins_turn_detection-1.0.0-py3-none-any.whl).

Installation

You can verify the build and install the generated wheel file directly into your environment:

pip install dist/livekit_plugins_turn_detection-1.0.0-py3-none-any.whl

Usage

Integrate the TurnDetectionModel into your LiveKit VoicePipelineAgent.

Basic Integration

from livekit.agents import JobContext, WorkerOptions, cli
from livekit.agents.voice import Agent, AgentSession
from livekit.plugins import openai, silero
from livekit.plugins.turn_detection import TurnDetectionModel

class Assistant(Agent):
    def __init__(self) -> None:
        super().__init__()

async def entrypoint(ctx: JobContext):
    await ctx.connect()

    session = AgentSession(
        vad=silero.VAD.load(),
        stt=openai.STT(language="vi"),
        llm=openai.LLM(),
        tts=openai.TTS(),
        # Initialize the turn detector
        turn_detection=TurnDetectionModel(threshold=0.5),
    )

    await session.start(
        agent=Assistant(),
        room=ctx.room
    )

if __name__ == "__main__":
    cli.run_app(WorkerOptions(entrypoint_fnc=entrypoint))

Configuration

The TurnDetectionModel accepts the following parameters:

• threshold (float, default 0.5): Probability threshold (0.0 - 1.0) to trigger end-of-turn. Higher values reduce false interruptions but may increase latency.