salmamohammedhamed22/arabic-eou-model

Arabic End-of-Utterance (EOU) Detection Model – Saudi Dialect

Base model: UBC-NLP/marbertv2 Model type: Sequence Classification (Binary: EOU vs CONTINUE) Language: Arabic (Saudi dialect focus) Trained on: 30,000 examples (≈15k positive + 15k negative) Framework: Hugging Face Transformers

---

Model Description

This model is a fine-tuned MarBERTv2 transformer for predicting End-of-Utterance (EOU) in Arabic conversations. It predicts whether a speaker has finished their turn based on transcription text, enabling real-time turn-taking in AI voice agents.

The model was trained specifically to integrate with LiveKit agents, which provide up to 6 previous turns to the EOU model. To leverage this, the training dataset includes sliding window context of up to 4 previous turns combined with the current utterance using the [SEP] token. This allows turn-aware predictions, rather than relying solely on the last utterance.

---

Training Data

• Source: Sada dataset (download here)
• Processed for EOU: Positive samples are full utterances; negative samples are incomplete prefixes (download here)
• Total samples: 30,000 (15k positive, 15k negative)
• Context: Up to 4 previous turns per example

---

Training Procedure

• Base model: UBC-NLP/marbertv2
• Fine-tuning task: Sequence classification (2 labels: EOU / CONTINUE)
• Labels:
  • 0: CONTINUE
  • 1: EOU
• Config updates:
  • hidden_dropout_prob = 0.3
  • attention_probs_dropout_prob = 0.3
• Training hyperparameters:
  • Learning rate: 3e-5
  • Weight decay: 0.05
  • Batch size: 32
  • Gradient accumulation: 4 steps
  • FP16 mixed precision: True
  • Epochs: 10
  • Warmup ratio: 0.1
  • Save strategy: per epoch, keep 1 best model
  • Evaluation metric for best model: eval_loss

---

Evaluation Results

Validation metrics (last epoch):

• Loss: 0.2556
• Accuracy: 0.9011
• F1: 0.8956
• Precision: 0.9488
• Recall: 0.8480

Training metrics (last epoch):

• Loss: 0.2025
• Accuracy: 0.9158
• F1: 0.9111
• Precision: 0.9649
• Recall: 0.8630

Test set results:

• Loss: 0.2624
• Accuracy: 0.8958
• F1: 0.8900
• Precision: 0.9419
• Recall: 0.8436

---

Intended Uses

Primary use-case

• LiveKit agents for real-time EOU detection
• Voice assistants
• Dialogue systems
• Turn-taking prediction in Arabic conversations

Not intended for

• Speech-to-text training
• General language modeling
• Speaker diarization

---

Limitations

• Primarily focused on Saudi dialect
• Negative prefixes may not capture all real-time partial ASR outputs
• Text-only model — does not process audio
• Performance may vary on other Arabic dialects

---

Inference Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "salmamohammedhamed22/arabic-eou-model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example input with context
text = "مرحبا كيف حالك [SEP] تمام الحمد لله."

inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
probs = torch.softmax(outputs.logits, dim=-1)
print(f"EOU probability: {probs[0][1].item():.3f}")  # probability of end-of-utterance