MohamedRashad/fleurs-ar-eg

Dataset Card for FLEURS Arabic–Egyptian Edition

Dataset Summary

FLEURS Arabic–Egyptian Edition is an unofficial, language-specific subset and adaptation of the FLEURS dataset, focused on Arabic (Egyptian) speech data. The dataset is designed for Automatic Speech Recognition (ASR) research and evaluation and follows the original FLEURS structure while being packaged as a standalone Arabic-focused dataset.

The data originates from the FLEURS (Few-shot Learning Evaluation of Universal Representations of Speech) corpus, which itself is the speech counterpart of the FLORES multilingual benchmark. This edition keeps the original sentence-level alignment, metadata, and splits, while standardizing audio to 16 kHz.

Dataset Structure

Data Instances

Each data sample consists of a read-speech audio recording paired with a sentence-level transcription. The utterances are derived from parallel sentences originally used in the FLORES benchmark.

Features

Feature	Type	Description
id	int32	Unique sample identifier
num_samples	int32	Number of audio samples
path	string	Path to the audio file
audio	audio (16 kHz)	Audio waveform
transcription	string	Normalized transcription
raw_transcription	string	Original (non-normalized) transcription
gender	class_label	Speaker gender (male, female, other)
lang_id	class_label	Language identifier (ar_eg)
language	string	Human-readable language name
lang_group_id	class_label	Language group classification

Language Groups

Arabic (ar_eg) belongs to the Central Asia / Middle East / North Africa (CMN) group, consistent with the original FLEURS taxonomy.

Dataset Splits

Split	Examples	Size
Train	2,104	~1.40 GB
Validation	295	~202 MB
Test	428	~300 MB
Total	2,827	~1.90 GB

How to Use

Load with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("MohamedRashad/fleurs-ar-eg", split="train")

Streaming Mode

from datasets import load_dataset

dataset = load_dataset(
    "MohamedRashad/fleurs-ar-eg",
    split="train",
    streaming=True
)

print(next(iter(dataset)))

PyTorch DataLoader

from datasets import load_dataset
from torch.utils.data import DataLoader

dataset = load_dataset("MohamedRashad/fleurs-ar-eg", split="train")
dataloader = DataLoader(dataset, batch_size=16, shuffle=True)

Dataset Creation

This dataset was created by:

1. Starting from the official FLEURS dataset
2. Selecting the Arabic (Egyptian) language configuration
3. Preserving the original sentence-level splits and metadata
4. Packaging the data into a standalone Hugging Face dataset
5. Keeping audio standardized at 16 kHz for ASR compatibility

No additional annotation, filtering, or transcription modification was performed beyond dataset restructuring.

Considerations for Using the Data

Social Impact

This dataset supports the development and evaluation of Arabic speech technologies, particularly for underrepresented dialects such as Egyptian Arabic. It aims to improve accessibility to speech-driven applications and multilingual research.

Biases

• The dataset consists of read speech, which may not reflect spontaneous conversational Arabic.
• Speaker demographics are limited and may not represent the full diversity of Egyptian Arabic speakers.
• Gender labels are coarse and self-reported or inferred.

Limitations

• Limited dataset size (~10 hours of audio)
• Domain restricted to read, formal sentences
• Accent variation within Egyptian Arabic is not explicitly labeled

License

This dataset is released under the Creative Commons Attribution (CC-BY) license, inherited from the original FLEURS dataset.

Users must provide appropriate attribution when using or redistributing the data.

Citation

If you use this dataset, please cite the original FLEURS paper:

@article{fleurs2022arxiv,
  title = {FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author = {Conneau, Alexis and Ma, Min and Khanuja, Simran and Zhang, Yu and Axelrod, Vera and Dalmia, Siddharth and Riesa, Jason and Rivera, Clara and Bapna, Ankur},
  journal = {arXiv preprint arXiv:2205.12446},
  url = {https://arxiv.org/abs/2205.12446},
  year = {2022}
}

Acknowledgements

Special thanks to the FLEURS authors and contributors for creating and releasing this multilingual speech benchmark, and to the Hugging Face ecosystem for tooling and support.