MohamedRashad/MGB-3-Arabic

Dataset Card for MGB-3 Arabic Speech Recognition

Dataset Summary

The MGB-3 Arabic dataset is a multi-genre collection of Egyptian Arabic speech extracted from YouTube videos, designed for speech recognition in challenging, real-world conditions. Unlike its predecessor MGB-2 which focused on broadcast TV news, MGB-3 emphasizes dialectal Arabic across diverse content types.

The dataset contains approximately 16 hours of Egyptian Arabic speech from 80 YouTube videos distributed across seven genres:

• Comedy
• Cooking
• Family/Kids
• Fashion
• Drama
• Sports
• Science (TEDx talks)

This dataset represents a low-resource, multi-genre scenario that challenges speech recognition systems to handle:

• Dialectal Arabic without well-defined orthographic systems
• Multiple speaking styles and acoustic conditions
• Spontaneous and scripted speech across different contexts

Supported Tasks

• Automatic Speech Recognition (ASR): The primary task is to transcribe Egyptian Arabic speech with time-stamped word recognition results

Languages

The dataset contains Egyptian Arabic (a dialect of Arabic, language code: ar).

Dataset Structure

Data Instances

Each instance contains:

• id: Unique identifier for the audio sample (string)
• audio: Audio file sampled at 16kHz
• text: Transcription of the speech (string)

Data Splits

The dataset is divided into three splits:

Split	Number of Examples	Size (bytes)
Train	2,153	~507 MB
Validation	1,927	~463 MB
Test	2,467	~597 MB
Total	6,547	~1.57 GB

Data Fields

• id (string): Unique identifier for each audio sample
• audio (Audio): Audio data with 16kHz sampling rate
  • array: The audio waveform as a numpy array
  • path: Path to the audio file
  • sampling_rate: 16000 Hz
• text (string): Ground truth transcription in Arabic

Dataset Creation

Curation Rationale

The MGB-3 challenge was created to address several research challenges:

1. Dialectal Arabic Recognition: Unlike Modern Standard Arabic (MSA), Egyptian Arabic lacks standardized orthography and has limited resources
2. Multi-genre Robustness: Testing systems across diverse content types and speaking styles
3. Low-resource Adaptation: Providing limited in-domain data (16 hours) to encourage adaptation techniques
4. Real-world Conditions: Using YouTube content represents "speech in the wild" with varying audio quality and spontaneous speech

Source Data

Initial Data Collection

The data was collected from YouTube videos across seven distinct genres, with videos carefully selected to represent Egyptian dialect. For the science category, TEDx talks in Egyptian dialect were chosen (talks up to 18 minutes in duration).

Who are the source language producers?

The speech comes from Egyptian Arabic speakers appearing in YouTube videos across various genres, including content creators, presenters, actors, speakers, and everyday people.

Annotations

Annotation process

The audio was manually transcribed with time-stamped alignments. The challenge emphasized dealing with the orthographic variability inherent in Egyptian Arabic transcription.

Who are the annotators?

The paper does not provide specific details about the annotators, but they were likely native Arabic speakers familiar with Egyptian dialect.

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to:

• Linguistic Diversity: Supporting speech technology for dialectal Arabic, particularly Egyptian Arabic
• Accessibility: Enabling better speech recognition for Arabic dialect speakers
• Research: Advancing multi-genre and low-resource speech recognition

Discussion of Biases

• The dataset focuses exclusively on Egyptian Arabic, which may not generalize to other Arabic dialects
• Genre distribution may not represent natural usage patterns
• YouTube content may have demographic biases in terms of speakers and topics

Other Known Limitations

• Limited Size: Only 16 hours of audio, making it challenging to train systems from scratch
• Orthographic Challenges: Egyptian Arabic lacks standardized spelling conventions
• Audio Quality: Variable quality due to YouTube source material
• Genre Balance: Equal distribution across genres may not reflect real-world usage

Additional Information

Dataset Curators

The dataset was created by researchers from:

• Qatar Computing Research Institute, HBKU, Doha, Qatar
• Centre for Speech Technology Research, University of Edinburgh, UK

Licensing Information

Please refer to the original MGB Challenge website for licensing information.

Citation Information

@article{ali2017speech,
  title={Speech Recognition Challenge in the Wild: Arabic MGB-3},
  author={Ali, Ahmed and Vogel, Stephan and Renals, Steve},
  journal={arXiv preprint arXiv:1709.07276},
  year={2017}
}