SoufianeDahimi/Tamazight-ASR-Dataset-v2

Tamazight-Arabic Speech Recognition Dataset

Dataset Description

This dataset contains speech segments in Tamazight (specifically focusing on the Tachelhit dialect) paired with their corresponding Arabic transcriptions. It is designed to support the development of automatic speech recognition (ASR) systems for the Tamazight language, particularly for translation into Modern Standard Arabic.

This is an actively growing dataset, with regular updates and new data points being added to improve coverage and quality.

Dataset Summary

• Total Examples: 20,344 audio segments
• Training Set: 18,309 examples
• Test Set: 2,035 examples
• Total Audio Duration: ~15.5 hours
• Languages: Tamazight (Tachelhit dialect) → Modern Standard Arabic
• Audio Format: WAV
• License: Apache 2.0
• Dataset Link: Hugging Face Dataset

Future Development

The dataset is actively expanding with plans to include other Tamazight dialects such as:

• Tarifit
• Central Atlas Tamazight
• Other regional variants

This ongoing development aims to create a comprehensive resource for Tamazight language technology.

Dataset Creation

Data Collection Process

The dataset was created through a multi-step pipeline:

1. Video Collection: Source videos were collected from YouTube, specifically targeting Tamazight movies with Arabic subtitles. The videos were downloaded using yt-dlp.

2. Subtitle Extraction: Currently using OCR services (via subtitleextractor.com) to extract hardcoded Arabic subtitles. This is an area where community contributions for open-source alternatives would be particularly valuable. If you have experience with OCR systems or know of better/faster/open-source solutions, your input would be greatly appreciated.

3. Audio-Text Synchronization: A custom processing pipeline was developed to:

   • Parse SRT files containing the extracted subtitles
   • Extract corresponding audio segments from the videos
   • Align the audio segments with their transcriptions
   • Generate a structured dataset with synchronized audio-text pairs

4. Dataset Preprocessing: Final preprocessing steps were applied to ensure data quality and consistency.

Data Structure

Each example in the dataset contains:

{
    'audio': {
        'path': str,        # Path to the audio file
        'array': np.array,  # The audio signal
        'sampling_rate': int # Sampling rate of the audio
    },
    'text': str,           # Arabic transcription
    'metadata': {
        'duration': float,  # Duration in seconds
        'start_time': float, # Start time in original video
        'end_time': float,  # End time in original video
        'language': str,    # Language code
        'source': str      # Source video identifier
    }
}

Dataset Statistics

• Training Split:

  • Number of examples: 18,309
  • Total size: ~8.9GB

• Test Split:

  • Number of examples: 2,035
  • Total size: ~992MB

Usage and Limitations

Intended Uses

This dataset is primarily intended for:

• Developing ASR systems for Tamazight (Tachlhit dialect)
• Training translation models from Tamazight speech to Arabic text
• Linguistic research on Tamazight-Arabic translation
• Preservation and digitization of Tamazight language content

Limitations and Biases

• Currently focuses only on the Tachlhit dialect of Tamazight
• Audio quality may vary due to source material
• Subtitle extraction process may introduce some transcription errors
• Dataset is biased towards movie content, which may not represent all speech contexts

Additional Information

Dataset Curators

This dataset is curated and maintained by Soufiane Dahimi. The project aims to create a comprehensive speech recognition resource for the Tamazight language, starting with the Tachelhit dialect and expanding to include other variants.

Citation

If you use this dataset in your research, please cite:

@dataset{tamazight_asr_2024,
  author       = {Soufiane Dahimi and contributors},
  title        = {Tamazight-Arabic Speech Recognition Dataset},
  year         = {2024},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SoufianeDahimi/Tamazight-ASR-Dataset-v2}}
}

Source Code

The complete data processing pipeline code will be released soon! This will include:

• Video downloading scripts
• Audio extraction and segmentation tools
• Subtitle processing utilities
• Dataset preparation notebooks

Stay tuned for the code release. Once available, community contributions to improve and extend the pipeline will be highly welcomed.

Contributions

This is a community-driven project aimed at creating a valuable resource for Tamazight language technology. There are several ways you can contribute:

1. Data Contributions

   • Add new audio-text pairs in Tachelhit dialect
   • Contribute data for other Tamazight dialects (Tarifit, Central Atlas, etc.)
   • Help validate and improve existing transcriptions

2. Technical Contributions

   • Suggest or implement better subtitle extraction methods
   • Improve the data processing pipeline
   • Share preprocessing techniques or model training approaches

3. Quality Improvements

   • Help with transcription verification
   • Report issues or inconsistencies
   • Suggest dataset enhancements

To contribute, you can:

• Open an issue or pull request on the dataset's repository
• Share your suggestions or improvements
• Reach out through Hugging Face's community features

All contributions will be acknowledged, and together we can build a more comprehensive resource for Tamazight language technology.

License

This dataset is licensed under Apache 2.0. Please see the LICENSE file for details.