🗣️ MasriSpeech-ASR-Finetuning: Egyptian Arabic Speech Fine-Tuning Dataset

🌍 Overview

MasriSpeech-ASR-Finetuning is a specialized subset of the MasriSpeech dataset, designed for fine-tuning Automatic Speech Recognition (ASR) models for Egyptian Arabic. This dataset contains 5,500 professionally annotated audio samples totaling over 100 hours of natural Egyptian Arabic speech.

💡 Key Features:

• High-quality 16kHz speech recordings
• Natural conversational Egyptian Arabic
• Speaker-balanced train/validation splits
• Comprehensive linguistic coverage
• Apache 2.0 license

📊 Dataset Summary

Feature	Value
Total Samples	5,500
Train Samples	5,000
Validation Samples	500
Sampling Rate	16 kHz
Total Duration	~100 hours
Languages	Egyptian Arabic (arz), Arabic (ar)
Format	Parquet
Dataset Size	1.19 GB
Download Size	1.13 GB
Annotations	Transcripts

🧱 Dataset Structure

The dataset follows Hugging Face datasets format with two splits:

DatasetDict({
    train: Dataset({
        features: ['audio', 'transcription'],
        num_rows: 5000
    })
    validation: Dataset({
        features: ['audio', 'transcription'],
        num_rows: 500
    })
})

Data Fields

• audio: Audio feature object containing:
  • Array: Raw speech waveform (1D float array)
  • Path: Relative audio path
  • Sampling_rate: 16,000 Hz
• transcription: string with Egyptian Arabic transcription

📈 Data Statistics

Split Distribution

Split	Examples	Size (GB)	Avg. Words	Empty	Non-Arabic
Train	5,000	1.11	13.34	0	0
Validation	500	0.08	9.60	0	0

Linguistic Analysis

Feature	Train Set	Validation Set
Top Words	في (2,025), و (1,698)	في (52), أنا (41)
Top Bigrams	(إن, أنا) (130)	(شاء, الله) (6)
Vocab Size	3,845	789
Unique Speakers	114	10

Word Count Distributions (Left: Train, Right: Validation)

How to Use ? 🧑‍💻

Loading with Hugging Face

from datasets import load_dataset
import IPython.display as ipd

# Load dataset (streaming recommended for large datasets)
ds = load_dataset('NightPrince/MasriSpeech-ASR-Finetuning', 
                 split='train',
                 streaming=True)

# Get first sample
sample = next(iter(ds))
print(f"Transcript: {sample['transcription']}")

# Play audio
ipd.Audio(sample['audio']['array'], 
          rate=sample['audio']['sampling_rate'])

Preprocessing the Dataset

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torch
model_name = "facebook/wav2vec2-base-960h"  # Spanish example
# or "facebook/wav2vec2-large-xlsr-53-en" for English
processor = Wav2Vec2Processor.from_pretrained(model_name)
model = Wav2Vec2ForCTC.from_pretrained(model_name)

def prepare_dataset(batch):
    audio = batch["audio"]
    
    # Extract audio array and sampling rate
    audio_array = audio["array"]
    sampling_rate = audio["sampling_rate"]
    
    # Process audio using feature extractor only
    inputs = processor.feature_extractor(
        audio_array, 
        sampling_rate=sampling_rate, 
        return_tensors="pt"
    )
    
    batch["input_values"] = inputs.input_values[0]
    
    # Process transcription using tokenizer only
    labels = processor.tokenizer(
        batch["transcription"], 
        return_tensors="pt"
    )
    
    batch["labels"] = labels["input_ids"][0]
    
    return batch

# Apply preprocessing to the entire dataset
print("Processing entire dataset...")
dataset = ds.map(prepare_dataset, remove_columns=["audio", "transcription"])

Fine-Tuning an ASR Model

from transformers import AutoModelForCTC, TrainingArguments, Trainer

# Load pre-trained model
model = AutoModelForCTC.from_pretrained("facebook/wav2vec2-base-960h")

# Define training arguments
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    num_train_epochs=3,
    save_steps=10,
    save_total_limit=2,
    logging_dir="./logs",
    logging_steps=10,
)

# Initialize Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
    eval_dataset=dataset,
)

# Train the model
trainer.train()

Evaluating the Model

# Evaluate the model
eval_results = trainer.evaluate()
print("Evaluation Results:", eval_results)

Exporting the Model

# Save the fine-tuned model
model.save_pretrained("./fine_tuned_model")
processor.save_pretrained("./fine_tuned_model")

📜 Citation

If you use MasriSpeech-ASR-Finetuning in your research or work, please cite it as follows:

@dataset{masrispeech_asr_finetuning,
  author       = {Yahya Muhammad Alnwsany},
  title        = {MasriSpeech-ASR-Finetuning: Egyptian Arabic Speech Fine-Tuning Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/collections/NightPrince/masrispeech-dataset-68594e59e46fd12c723f1544}
}

📜 Licensing

This dataset is released under the Apache 2.0 License. You are free to use, modify, and distribute the dataset, provided you comply with the terms of the license. For more details, see the LICENSE.

🙌 Acknowledgments

We would like to thank the following for their contributions and support:

• Annotators: For their meticulous work in creating high-quality transcriptions.
• Hugging Face: For providing tools and hosting the dataset.
• Open-Source Community: For their continuous support and feedback.

💡 Use Cases

MasriSpeech-ASR-Finetuning can be used in various applications, including:

• Fine-tuning Automatic Speech Recognition (ASR) models for Egyptian Arabic.
• Dialectal Arabic linguistic research.
• Speech synthesis and voice cloning.
• Training and benchmarking machine learning models for low-resource languages.

🤝 Contributing

We welcome contributions to improve MasriSpeech-ASR-Finetuning. If you have suggestions, find issues, or want to add new features, please:

1. Fork the repository.
2. Create a new branch for your changes.
3. Submit a pull request with a detailed description of your changes.

For questions or feedback, feel free to contact the maintainer.

📝 Changelog

[1.0.0] - 2025-08-02

• Initial release of MasriSpeech-ASR-Finetuning.
• Includes 5,500 audio samples with transcriptions.
• Train/validation splits provided.
• Dataset hosted on Hugging Face.