anassdabaghi/tts_darija

language:

• —ar license: cc-by-4.0 task_categories:
• —automatic-speech-recognition task_ids:
• —automatic-speech-recognition prettyname: Darija Arabic Speech Dataset sizecategories:
• —1K<n<10K tags:
• —darija
• —moroccan-arabic
• —arabic
• —speech
• —asr
• —automatic-speech-recognition
• —whisper
• —morocco ---

Moroccan Darija Speech Dataset

A speech dataset for Moroccan Arabic (Darija) automatic speech recognition (ASR).

The dataset consists of short audio clips extracted from publicly available videos, paired with their corresponding Arabic/Darija transcriptions obtained from subtitle files.

Dataset Summary

This dataset is designed primarily for training and evaluating automatic speech recognition models for Moroccan Darija.

Dataset Structure

Each example contains:

audio
text

Example:

{
    "audio": {
        "array": [...],
        "sampling_rate": 16000
    },
    "text": "..."
}

The audio column contains the speech recording and the text column contains its corresponding transcription.

Data Collection

The audio data was collected from publicly available online videos.

Subtitle files were used to identify speech segments and their corresponding transcriptions. Each subtitle cue was converted into an individual WAV audio segment using FFmpeg.

The processing pipeline:

1. 1.Download audio from online videos.
2. 2.Obtain Arabic subtitle files.
3. 3.Parse subtitle timestamps.
4. 4.Match each subtitle cue with its corresponding audio segment.
5. 5.Extract the segment using FFmpeg.
6. 6.Convert audio to mono WAV at 16 kHz.
7. 7.Remove segments shorter than 0.3 seconds.
8. 8.Remove segments longer than 20 seconds.
9. 9.Clean the transcription text.
10. 10.Store each audio/transcription pair as a dataset example.

Preprocessing

Audio is normalized to:

• —Sampling rate: 16,000 Hz
• —Channels: 1 (mono)
• —Format: WAV

Subtitle text is normalized by:

• —Removing excessive whitespace.
• —Removing selected punctuation characters.
• —Keeping Arabic characters and the original transcription content.

Intended Use

This dataset is intended for research and experimentation in:

• —Moroccan Darija ASR
• —Arabic speech recognition
• —Whisper fine-tuning
• —Speech-to-text systems
• —Low-resource language ASR
• —Moroccan Arabic NLP

Example models that could potentially be fine-tuned using this dataset include:

• —Whisper
• —wav2vec 2.0
• —XLS-R
• —HuBERT
• —other Transformer-based ASR models

Limitations

The dataset has several limitations:

• —Transcriptions originate from subtitles and may contain errors.
• —Subtitle timing may not perfectly match the spoken audio.
• —Some subtitles may contain spelling inconsistencies or dialect variations.
• —The dataset may contain different speakers, recording environments, accents, and audio qualities.
• —The dataset has not necessarily been manually validated at the individual clip level.
• —Some clips may contain non-speech segments or background noise.

For high-quality ASR training, additional filtering and manual validation are recommended.

Recommended Data Splitting

When creating training, validation, and test sets, splitting should preferably be performed at the video level rather than the clip level.

Multiple clips can originate from the same video. Randomly splitting individual clips could therefore result in clips from the same source appearing in both training and evaluation sets.

Citation

If you use this dataset in your research or project, please cite the dataset repository:

@misc{dabaghi_darija_asr,
  author = {Dabaghi, Anass},
  title = {Moroccan Darija Speech Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/Dabaghi-Anass/darija-asr-dataset}
}

License and Data Rights

The underlying audio and subtitle material may be subject to the rights, licenses, and terms applicable to their original sources.

This dataset is provided for research and educational purposes. Users are responsible for verifying that their intended use complies with the applicable copyright, licensing, privacy, and platform terms for the original material.

Contact

Created by Anass Dabaghi.