MARBERT Arabic End-of-Utterance Detection
Fine-tuned MARBERT model for Arabic End-of-Utterance (EOU) detection in real-time voice agents.
Model Description
- Base Model: UBC-NLP/MARBERT (163M parameters)
- Task: Binary sequence classification (complete vs incomplete utterance)
- Language: Arabic (emphasis on Saudi/Gulf dialect)
- Training Data: 125K samples from SADA22 dataset
- Inference Speed: ~30ms average latency on CPU
Performance
| Metric | Score |
|---|---|
| F1 Score | 0.8174 |
| Accuracy | 0.7995 |
| Precision | 0.7506 |
| Recall | 0.8971 |
| AUC-ROC | 0.8249 |
Test Set: 31,289 samples (50% complete, 50% incomplete)
Usage
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model = AutoModelForSequenceClassification.from_pretrained("azeddinShr/marbert-arabic-eou")
tokenizer = AutoTokenizer.from_pretrained("azeddinShr/marbert-arabic-eou")
def predict_eou(text):
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
outputs = model(**inputs)
probs = torch.softmax(outputs.logits, dim=-1)
eou_prob = probs[0][1].item()
return eou_prob
# Example
text = "شكرا جزيلا على المساعدة"
prob = predict_eou(text)
is_complete = prob > 0.5
print(f"EOU Probability: {prob:.3f} - {'Complete' if is_complete else 'Incomplete'}")
Training Details
- Epochs: 6
- Batch Size: 16 (train), 32 (eval)
- Learning Rate: 2e-5
- Optimizer: AdamW
- Max Length: 128 tokens
- Training Time: ~2 minutes (GPU)
Use Cases
- Real-time Arabic voice agents
- Turn-taking detection in conversations
- Streaming speech-to-text applications
- Voice assistant interrupt handling
Limitations
- Best performance on Saudi/Gulf Arabic dialects
- Requires Arabic text input (not audio)
Citation
@model{marbert-arabic-eou,
author = {azeddinShr},
title = {MARBERT Arabic End-of-Utterance Detection},
year = {2025},
publisher = {HuggingFace},
url = {https://huggingface.co/azeddinShr/marbert-arabic-eou}
}
Dataset
Training dataset: azeddinShr/arabic-eou-sada22