MadLook/arabic-whisper-multidialect

Arabic Whisper Multi-Dialect ASR Dataset

A comprehensive multi-dialect Arabic speech recognition dataset prepared for Whisper model fine-tuning.

Dataset Description

This dataset combines high-quality Arabic speech data from multiple dialects, specifically curated for fine-tuning OpenAI's Whisper models on Arabic speech recognition tasks.

Dialects Included

• Modern Standard Arabic (MSA) - Formal Arabic used in media and formal contexts
• Egyptian Arabic (EGY) - The most widely understood Arabic dialect
• Gulf Arabic (GULF) - Arabic dialects from the Gulf region

Dataset Statistics

• Total Samples: ~105,000 audio-text pairs
• Train Split: ~94,500 samples (90%)
• Validation Split: ~5,250 samples (5%)
• Test Split: ~5,250 samples (5%)
• Audio Format: 16kHz, mono
• Text: Normalized Arabic script

Source Datasets

This dataset is a combination of the following sources:

1. MightyStudent/Egyptian-ASR-MGB-3 - Egyptian dialect broadcast data
2. MAdel121/arabic-egy-cleaned - Cleaned Egyptian Arabic speech
3. Nourhann/filtered_common_voices_16khz - Modern Standard Arabic
4. badrex/arabic-speech-SADA22-Khaliji - Gulf dialect speech (30% sample)

Dataset Structure

DatasetDict({
    train: Dataset({
        features: ['audio', 'sentence'],
        num_rows: ~94500
    })
    validation: Dataset({
        features: ['audio', 'sentence'],
        num_rows: ~5250
    })
    test: Dataset({
        features: ['audio', 'sentence'],
        num_rows: ~5250
    })
})

Data Fields

• audio: Audio file (16kHz sampling rate)
• sentence: Transcription text in Arabic
• dialect: Dialect label ('msa', 'egy', or 'gulf')

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("MadLook/arabic-whisper-multidialect")

print(dataset)
# Access splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

Fine-tuning Whisper

from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Arabic", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Prepare dataset
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor.feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
    batch["labels"] = processor.tokenizer(batch["sentence"]).input_ids
    return batch

dataset = dataset.map(prepare_dataset)

# Continue with training...

Data Processing

The dataset has been processed with the following steps:

1. Audio normalization: All audio resampled to 16kHz mono
2. Text normalization: Arabic text normalized (diacritics removed, Alef forms unified)
3. Dialect balancing: Datasets sampled to ensure balanced representation
4. Quality filtering: Low-quality samples removed
5. Train/Val/Test split: 90/5/5 split with shuffling

Intended Use

Primary Use Cases

• Fine-tuning Whisper models for Arabic ASR
• Multi-dialect Arabic speech recognition research
• Arabic speech technology development
• Dialect-aware speech recognition systems

Considerations

• This dataset combines multiple dialects; consider dialect-specific fine-tuning if needed
• Text is normalized and may not preserve dialectal spelling variations
• Dataset is balanced across dialects but individual dialect performance may vary

Limitations

• Gulf dialect representation is lower (30% of original dataset)
• No code-switching samples included
• Limited to broadcast and read speech domains
• May not generalize well to conversational or noisy speech

Citation

If you use this dataset, please cite the original source datasets:

@dataset{arabic_whisper_multidialect,
  title={Arabic Whisper Multi-Dialect Dataset},
  author={MadLook},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/MadLook/arabic-whisper-multidialect}}
}

License

This dataset combines data from multiple sources. Please refer to the original datasets for their respective licenses:

• MightyStudent/Egyptian-ASR-MGB-3
• MAdel121/arabic-egy-cleaned
• Nourhann/filtered_common_voices_16khz
• badrex/arabic-speech-SADA22-Khaliji

Acknowledgments

This dataset builds upon the excellent work of the Arabic NLP and speech recognition community. Special thanks to the creators of the source datasets for making their data publicly available.