README.md

metadata

language: ar
license: mit
tags:
  - end-of-utterance
  - dialogue
  - text-classification
dataset_size: 30000
task_categories:
  - text-classification
task_ids:
  - multi-class-classification

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

Dataset type: Binary classification (EOU vs non-EOU) Language: Arabic (Saudi dialect focus) Size: 30,000 examples (≈15k positive + 15k negative) Format: JSON list of { "text": ..., "label": ... }

---

Dataset Summary

This dataset contains conversational Arabic utterances labeled for End-of-Utterance (EOU) detection. The goal is to train models that can predict whether a speaker has finished speaking based on transcription text only, enabling real-time turn-taking in voice agents.

The dataset is especially tailored to Saudi dialect Arabic, but also includes general conversational Arabic.

It contains:

• Positive samples (label = 1): Full utterances representing completed turns.
• Negative samples (label = 0): Incomplete prefixes generated from each utterance to simulate non-final turns.

LiveKit Context Note

LiveKit provides up to 6 previous turns to the EOU model for prediction. To take advantage of this, each example includes sliding window conversational context using up to 4 previous turns combined with the target utterance using the token [SEP]. This allows the model to learn turn-aware EOU detection rather than only relying on the last utterance.

---

Dataset Structure

Each JSON record looks like:

{
  "text": "مرحبا كيف حالك [SEP] تمام الحمد لله",
  "label": 1
}

• text: string — full utterance or context + utterance joined with [SEP]
• label:
  • 1 → End of utterance
  • 0 → Not end of utterance (incomplete prefix)

---

How the Dataset Was Created

We used a custom Python script (included in the repo) to generate both positive and negative examples from the original ASR-aligned CSV transcripts from the Sada dataset (download here the train.csv).

1. Preprocessing

• Remove excessive spacing
• Preserve Arabic and English punctuation
• Sort utterances by timestamp
• Group by audio file name

2. Sliding Window Context

For each utterance, we generate multiple samples with context length:

0, 1, 2, 3, 4 previous turns

Example:

U1 [SEP] U2 [SEP] U3

This context design aligns with LiveKit’s turn input (up to 6 turns).

3. Positive Samples (label = 1)

Each full utterance and its context version becomes a positive example.

4. Negative Samples (label = 0)

We generate up to 5 incomplete prefixes per utterance:

Example:

"انا كنت" → 0  
"انا كنت أبغى" → 0  
"انا كنت أبغى اسوي" → 0  

Prefixes simulate real-time partial speech coming from an STT system.

5. Balancing

• 15k positives
• 15k negatives
• Total: 30,000 examples

6. Final Shuffle

Dataset is shuffled globally to prevent ordering effects.

---

Intended Uses

Primary use-case

Training end-of-utterance detection models for:

• LiveKit agents
• Voice assistants
• Real-time dialog systems
• Arabic conversational AI
• Turn-taking prediction

Not intended for

• STT training
• Language modeling
• Speaker diarization

---

Limitations

• Primarily focused on Saudi dialect
• Generated negative prefixes may not capture all real-time ASR errors
• No audio included (text-only dataset)