atlasia/DODa-audio-dataset-V3

Moroccan Darija Speech Dataset

Overview

This dataset consists of 12,743 parallel text and speech samples for Moroccan Darija, including its transcription in both Latin and Arabic scripts and English translations. It was created to support speech recognition, language modeling, and NLP tasks for Moroccan Darija.

Dataset Source

The dataset was originally sourced from this repository, where it was available as a CSV file containing three columns:

• darija: Sentences in Moroccan Darija using Latin letters.
• eng: English translations of the sentences.
• darija_ar: Sentences in Moroccan Darija using Arabic script.

Data Preprocessing

To ensure data quality, the following preprocessing steps were applied:

1. Removed all rows with missing values in any of the three original columns.
2. Filtered dataset to retain only fully filled rows.
3. The cleaned dataset resulted in 12,743 sentences ready for audio recording.

Audio Recording Process

The dataset was recorded by 7 contributors (4 females, 3 males). Each sentence was spoken and recorded by one of the contributors. The dataset was divided into 13 chunks, but contributors were not assigned strictly per chunk—some recorded more than 1,000 sentences, while others recorded fewer.

Recording Distribution

0-999     -> Female
1000-1999 -> Male
2000-2730 -> Female
2731-2800 -> Male
2801-2999 -> Male
3000-3999 -> Male
4000-4999 -> Male
5000-5999 -> Female
6000-6999 -> Male
7000-7999 -> Female
8000-8999 -> Female
9000-9999 -> Male
10000-10999 -> Male
11000-11999 -> Male
12000-12350 -> Male
12351-12742 -> Male

• The recorded audio files were standardized to 16kHz sample rate to maintain consistency.

Transcription Correction Process

During the dataset review, spelling errors were identified in the darija_ar (Arabic script) column. For example:

• Incorrect: "شوكران"
• Correct: "شكرا"

Correction Methodology

To correct these errors efficiently:

1. We used Wit.ai (Arabic language setting) to transcribe all audio recordings.
2. Wit.ai skipped/ignored words that were not in Arabic dialects.
3. Instead of manually correcting every sentence, we added the missing words to the transcription.
4. This method ensured high accuracy and eliminated human error in transcription.

Final Dataset Structure

After full correction and validation, the final dataset consists of five columns:

Column Name	Description
audio	Speech recordings for Darija sentences
darija_Ltn	Darija sentences using Latin letters
darija_Arab_new	Corrected Darija sentences using Arabic script
english	English translation of Darija sentences
darija_Arab_old	Original (uncorrected) Darija sentences in Arabic script

Applications

This dataset can be used for:

• Speech-to-text models for Moroccan Darija.
• NLP applications, such as machine translation and text-to-speech synthesis.
• Linguistic studies on Moroccan Darija.
• Automatic pronunciation analysis.

Acknowledgments

Contact

For any inquiries or collaborations, feel free to connect with me on LinkedIn. We would like to extend special thanks to the contributors:

• BIDRY Mahmoud
• ZAIDOUNE Youssef
• AtlasIA's community.

License

This dataset is released under the MIT License. Feel free to use, modify, and distribute it for research and development purposes.

Citation

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

@misc{darija_speech_dataset,
  author = {BIDRY Mahmoud, ZAIDOUNE Youssef, et al.},
  title = {Moroccan Darija Speech Dataset},
  year = {2025},
  organization = {atlasIA}
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/atlasia/DODa-audio-dataset-V3}
}