mahmoudsaalama/arabic-eou-camelbert

Arabic End-of-Utterance Detection Model

Fine-tuned CAMeLBERT model for detecting end-of-utterance in Arabic conversations, with emphasis on Saudi dialect.

Model Description

This model is designed to detect when a speaker has finished their conversational turn in Arabic dialogue. It's particularly optimized for Saudi dialect patterns and real-time applications.

Model Details

• Base Model: CAMeLBERT-MSA (CAMeL-Lab/bert-base-arabic-camelbert-msa)
• Task: Binary classification (EOU vs. non-EOU)
• Language: Arabic (Modern Standard Arabic + Saudi dialect)
• Parameters: ~110M (base encoder) + classification head
• Training Data: 2,000+ Arabic conversation samples

Intended Use

• Real-time turn detection in conversational AI agents
• Voice assistants for Arabic speakers
• Dialogue systems
• LiveKit agent integration

How to Use

Installation

pip install torch transformers

Basic Usage

from transformers import AutoTokenizer, AutoModel
import torch

# Load model and tokenizer
model_name = "mahmoudsaalama/arabic-eou-camelbert"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

# Prepare input
text = "السلام عليكم ورحمة الله"
inputs = tokenizer(text, return_tensors="pt", max_length=128, truncation=True)

# Get prediction
with torch.no_grad():
    outputs = model(**inputs)
    probability = torch.sigmoid(outputs.logits).item()
    is_eou = probability > 0.5

print(f"EOU Probability: {probability:.4f}")
print(f"Is EOU: {is_eou}")

Using the SDK

For easier integration, use the Arabic EOU SDK:

pip install arabic-eou-sdk

from arabic_eou_sdk import ArabicEOUDetector

detector = ArabicEOUDetector(model_name="mahmoudsaalama/arabic-eou-camelbert")
result = detector.update_transcription("السلام عليكم", is_final=True)

print(f"Is EOU: {result['is_eou']}")
print(f"Probability: {result['probability']:.4f}")
print(f"Confidence: {result['confidence']:.4f}")

Training Details

Training Data

• Size: ~2,000 samples (1,600 train, 200 val, 200 test)
• Balance: 50% positive (EOU), 50% negative (non-EOU)
• Sources: Synthetic Saudi Arabic conversations + public Arabic datasets

Training Procedure

• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 16
• Epochs: 10 (with early stopping)
• Mixed Precision: FP16
• Hardware: GPU (CUDA)

Evaluation Metrics

Metric	Score
Accuracy	~90%
Precision	~88%
Recall	~92%
F1 Score	~90%
ROC AUC	~95%

Inference Speed

Configuration	Latency
GPU (FP32)	~15-20ms
GPU (INT8)	~8-12ms
CPU (FP32)	~60-80ms
CPU (INT8)	~25-35ms

Limitations

• Dialectal Coverage: Optimized for Saudi dialect, may not generalize perfectly to other Arabic dialects
• Synthetic Data: Trained primarily on synthetic conversations
• Domain: Limited to common conversational topics
• Dataset Size: Relatively small training set

Bias and Fairness

• Model may perform better on Saudi dialect than other Arabic dialects
• Training data focuses on common conversational patterns
• May not handle code-switching or mixed-language conversations well

Citation

@model{arabic_eou_camelbert_2025,
  author = {Mahmoud Saalama},
  title = {Arabic End-of-Utterance Detection Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/mahmoudsaalama/arabic-eou-camelbert}
}

License

MIT License

Contact

For questions or feedback:

• GitHub: arabic-eou-livekit
• Hugging Face: @mahmoudsaalama

Acknowledgments

• Base model: CAMeLBERT by CAMeL Lab
• Framework: Transformers by Hugging Face
• Integration: LiveKit for real-time applications