livekit-plugins-marbert 0.1.0

Arabic End-of-Utterance detection for LiveKit using MARBERT

LiveKit MARBERT Turn Detector

Arabic End-of-Utterance detection plugin for LiveKit Agents using fine-tuned MARBERT.

Installation

pip install livekit-plugins-marbert

Or install from source:

pip install git+https://github.com/azeddinshr/livekit-plugins-marbert.git

Quick Start

from livekit.plugins.marbert import MarbertTurnDetector
from livekit.agents import Agent, AgentSession, JobContext, JobProcess, WorkerOptions, cli
from livekit.plugins import openai, silero

def prewarm(proc: JobProcess):
    """Load models once at startup (recommended for production)."""
    proc.userdata["vad"] = silero.VAD.load()
    proc.userdata["marbert"] = MarbertTurnDetector(
        model_name="azeddinShr/marbert-arabic-eou",
        device="cpu"
    )

async def entrypoint(ctx: JobContext):
    await ctx.connect()
    
    session = AgentSession(
        stt=openai.STT(model="gpt-4o-transcribe", language="ar"),
        llm=openai.LLM(model="gpt-4o-mini"),
        tts=openai.TTS(model="gpt-4o-mini-tts"),
        vad=ctx.proc.userdata["vad"],
        turn_detection=ctx.proc.userdata["marbert"]
    )
    
    agent = Agent(instructions="أنت مساعد سعودي ذكي")
    await session.start(agent=agent, room=ctx.room)

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

See examples/basic_usage.py for a complete working example.

Features

• ✅ Fine-tuned on 125K Arabic utterances (SADA22 dataset)
• ✅ Specialized for Saudi dialect
• ✅ 77% F1 score, 30ms average CPU latency
• ✅ Drop-in replacement for LiveKit turn detector
• ✅ Compatible with LiveKit Agents 1.3.9+
• ✅ Production-ready with prewarm support

Performance

Metric	MARBERT	LiveKit Multilingual
F1 Score	0.77	0.66*
Latency (CPU)	30ms	300ms*
Language	Arabic (Saudi)	14 languages
Model Size	163M params	500M params

*Comparison with community Arabic model

Model Details

• Base Model: UBC-NLP/MARBERT (163M parameters)
• Fine-tuned Model: azeddinShr/marbert-arabic-eou
• Training Dataset: azeddinShr/arabic-eou-sada22
• Task: Binary classification (complete vs incomplete utterance)
• Training Data: 125K samples from SADA22 (Saudi dialect focus)

Configuration

MarbertTurnDetector(
    model_name="azeddinShr/marbert-arabic-eou",  # HuggingFace model ID
    threshold=0.5,                                 # EOU probability threshold
    device="cpu"                                   # "cpu" or "cuda"
)

Requirements

• Python >= 3.9
• livekit-agents >= 1.3.9
• transformers >= 4.30.0
• torch >= 2.0.0

License

Apache 2.0

Citation

If you use this model in your research, please cite:

@misc{marbert-arabic-eou,
  author = {Sahir, Azeddin},
  title = {MARBERT Arabic End-of-Utterance Detection},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/azeddinshr/livekit-plugins-marbert}
}

Links

• GitHub Repository
• Fine-tuned Model
• Training Dataset
• LiveKit Agents Documentation