arabic-eou-model

Model Summary

arabic-eou-turn-taker is a lightweight transformer model fine-tuned for Arabic End-of-Utterance (EOU) detection in real-time conversational speech. It predicts whether the speaker has finished their turn (END) or is still speaking (CONTINUE).

Designed specifically for:

•	Arabic conversational voice agents
•	LiveKit turn-taking modules
•	Real-time ASR pipelines
•	Saudi/Gulf dialects support

Intended Use

This model should be used in:

•	Voice assistants
•	Customer support agents
•	Streaming ASR applications
•	Real-time AI call agents
•	LiveKit agents requiring turn-taking

Model Details

•	Architecture: Transformer-based sequence classifier
•	Labels:
•	0 = CONTINUE
•	1 = END
•	Languages: Arabic (MSA + dialects(mostly saudi))
•	Tokenizer: AutoTokenizer (loaded automatically)

Training Dataset

dataset: robamahmoudd/arabic-eou-balanced

•	train.jsonl → 106,920 samples
•	test.jsonl → 26,731 samples
•	Sources: conversational Arabic, MADAR dialect corpus, SADA22 Arabic Dataset

Labeling Guidelines

•	END if utterance naturally ends or ends with ؟ . !
•	CONTINUE if utterance is incomplete, introductory, or cut off

Evaluation Results

Accuracy ~0.994

F1 Score ~0.994

Precision ~1.00

Recall ~0.98

Inference Example

from transformers import pipeline

clf = pipeline("text-classification", model="YOUR_USERNAME/arabic-eou-turn-taker", return_all_scores=True)

clf("ده طبيعي ولا لا؟")

or using SDK:

from eou_detector import ArabicEOUDetector
detector = ArabicEOUDetector("YOUR_USERNAME/arabic-eou-turn-taker")
detector.is_end("ده طبيعي ولا لا؟")