metadata
language: ar
base_model: faisalq/SaudiBERT
tags:
- arabic
- saudi
- eou
- turn-taking
- conversational-ai
license: mit
Saudi Arabic End-of-Utterance (EOU) Model
This model detects End-of-Utterance (EOU) events in Saudi Arabic conversational text.
It outputs the probability that a speaker has finished their turn, enabling natural turn-taking in real-time voice agents (e.g., LiveKit).
Task
Binary classification (probability output):
- 0 → Incomplete utterance (speaker likely to continue)
- 1 → Complete utterance (end of turn)
Model Details
- Base model:
faisalq/SaudiBERT - Architecture: BERT Sequence Classification
- Output: Single probability (sigmoid)
- Dialect focus: Saudi Arabic (ar-SA)
Training
- Dataset: Saudi Arabic conversational EOU dataset
https://huggingface.co/datasets/HussainKAUST/saudi-data-eou.jsonl - Data source: Synthetic Saudi dialogue with natural pauses and incomplete turns
- Loss: Focal Loss (class imbalance handling)
- Epochs: 6