Hishambarakat/Bahraini_Speech_Dataset

Bahraini Speech Dataset

Overview

Bahraini Speech Dataset is a Bahraini Arabic speech corpus built from publicly available podcast/video content and processed into single-speaker utterance clips with aligned transcriptions.

The dataset is designed to support research and experimentation in:

• Automatic Speech Recognition (ASR)
• Dialectal Arabic modeling
• Low-resource speech + language workflows
• Phonetic / linguistic analysis

Key Properties

• Total clips: 90,421
• Train / Test split: 87,708 / 2,713
• Split method: Stratified by speaker gender
• Language / Dialect: Arabic (Bahraini dialect; may include code-switching)
• Audio format: WAV, mono
• Sampling rate: 24,000 Hz
• Typical clip duration: ~1–6 seconds
• Approx. total audio duration: ~50+ hours
• Speakers: Dozens of anonymized speaker IDs (male / female / unknown)

Quick Start (Loading)

import os
from pathlib import Path
from huggingface_hub import snapshot_download
from datasets import load_from_disk

REPO_ID = "Hishambarakat/Bahraini_Speech_Dataset"

local_repo = snapshot_download(
    repo_id=REPO_ID,
    repo_type="dataset",
    allow_patterns=[
        "dataset_dict.json",
        "dataset_info.json",
        "train/**",
        "test/**",
    ],
)

ds = load_from_disk(local_repo)
print(ds)
print(ds["train"][0].keys())

Dataset Contents

Each sample in the dataset contains:

• Audio clip (WAV, mono, 24 kHz)
• Transcription text (Arabic, may include code-switching)
• Speaker ID (anonymized, non-identifying)
• Speaker gender (male / female / unknown)
• Clip duration
• Sampling rate

The dataset is split into:

• train
• test

using a stratified split by speaker gender.

---

Data Source & Collection Method

• Audio is derived from publicly available online content, such as podcasts and videos published openly on platforms like YouTube.
• The content was programmatically segmented into single-speaker utterances using speaker diarization and alignment pipelines.
• No private, paid, or restricted data sources were used.
• Speaker identities are not known and not recoverable from this dataset.

---

Pipeline Process Overview

This dataset was built using a multi-stage pipeline that processes raw audio into clean, transcribed speech clips:

Stage 1: Audio Segmentation

Raw YouTube videos are downloaded and split into 10-minute segments with intro/outro trimming using ffmpeg.

Stage 2: Speaker Diarization

Each segment is processed to identify who speaks when using:

• pyannote/speaker-diarization-3.1 - State-of-the-art neural speaker diarization
• Outputs RTTM (Rich Transcription Time Marked) files with speaker timestamps

Stage 3: Speaker Sample Extraction & Gender Detection

• Extracts clean single-speaker clips for each speaker
• Auto-detects speaker gender using pitch analysis (librosa)
• Creates sample audio for speaker verification

Stage 4: Global Speaker Linking

Local speaker IDs (per-segment) are linked into consistent global speaker IDs across all content using:

• speechbrain/spkrec-ecapa-voxceleb - ECAPA-TDNN speaker embeddings for voice matching
• Agglomerative clustering (cosine distance) to identify the same speaker across different segments

Stage 5: Clip Dataset Creation

Single-speaker intervals are extracted with quality filtering:

• faster-whisper (large-v3) - ASR purity check to filter out unclear speech
• Removes clips with background noise, overlapping speakers, or low confidence scores
• Adds padding and splits long intervals into 1-6 second clips

Stage 6: Batch Transcription

All clips are transcribed using:

• MohamedRashad/Arabic-Whisper-CodeSwitching-Edition - Arabic/English code-switching ASR model
• Batch processing with chunking for efficient GPU utilization
• Extracts text, language, and confidence scores

Stage 7: Dataset Building

Final Hugging Face dataset is built with:

• Stratified train/test split by speaker gender
• 24kHz audio resampling for TTS compatibility
• Arrow format for efficient loading and streaming

---

Intended Use

This dataset is intended for:

• Academic research
• Non-commercial experimentation
• Speech and language model training (research only)
• Dialect analysis

❌ Not Intended For

• Commercial speech products
• Voice cloning of identifiable individuals
• Any use that violates the original content creators’ rights

If you wish to use this dataset for commercial purposes, you are responsible for ensuring that you have the appropriate rights from the original content owners.

---

Ethical & Legal Notice

• This dataset is provided without any warranty.
• The dataset creator does not claim ownership of the original audio content.
• Responsibility for downstream usage lies entirely with the user.
• If any content owner believes material should be removed, please open an issue or contact the dataset author.

---

File Structure

bahrain_podcast_hf_export/
├── asr_dataset/
│   ├── train/
│   ├── test/
│   ├── dataset_dict.json
│   └── state.json
└── README.md

Audio paths inside the dataset are relative, making the dataset portable across machines once downloaded from Hugging Face.

---

Acknowledgments & Credits

This dataset was built using several open-source tools and models. We gratefully acknowledge:

Core Technologies

• pyannote.audio - Speaker diarization toolkit

  • Model: pyannote/speaker-diarization-3.1

• SpeechBrain - Speaker recognition and embeddings

  • Model: speechbrain/spkrec-ecapa-voxceleb

• faster-whisper - Efficient Whisper inference for quality filtering

  • Model: Whisper large-v3

• Arabic-Whisper-CodeSwitching-Edition - Arabic/English ASR model by Mohamed Rashad

• librosa - Audio analysis and pitch detection

• FFmpeg - Audio processing and segmentation

• Hugging Face Datasets - Dataset management and distribution

References

@inproceedings{Bredin23,
  author={Hervé Bredin},
  title={{pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe}},
  year=2023,
  booktitle={Proc. INTERSPEECH 2023},
}

@misc{speechbrain,
  title={{SpeechBrain: A General-Purpose Speech Toolkit}},
  author={Mirco Ravanelli and Titouan Parcollet and others},
  year={2021},
  publisher={GitHub},
  url={https://github.com/speechbrain/speechbrain}
}

@misc{rashad2024arabicwhisper,
  title={Arabic-Whisper-CodeSwitching-Edition},
  author={Mohamed Rashad},
  year={2024},
  url={https://huggingface.co/spaces/MohamedRashad/Arabic-Whisper-CodeSwitching-Edition},
}

---

Citation

If you use this dataset in your work, please cite it as follows:

@dataset{barakat_bahraini_speech_2026,
  author       = {Hisham Barakat},
  title        = {Bahraini Speech Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Hishambarakat/Bahraini_Speech_Dataset},
  LinkedIn     = {https://www.linkedin.com/in/hishambarakat/}
}