MoulSot-Full Dataset

Dataset Summary

MoulSot-Full is a large-scale Moroccan Darija speech dataset containing in total 1,500 hours of speech audio. From this extensive corpus, a high-quality subset of approximately 80 hours has been carefully curated and transcribed. It was built entirely from publicly available YouTube content across 51 diverse channels (including vlogs, podcasts, interviews, and commentary) to capture real-world Moroccan Darija, including natural code-switching with French, various regional dialects, and diverse acoustic conditions.

The full 1,500-hour dataset was processed using an automated pipeline involving:

1. Voice Activity Detection: Silero VAD.
2. Audio Quality Scoring: SQUIM (PESQ, STOI) and Meta's Audiobox Aesthetics.
3. Speech Enhancement: DNS64 for low-quality segments.
4. Transcription: Automated annotation using Gemini 2.5 Pro with specific prompting to preserve code-switched Latin script inside the Arabic transcriptions.

This dataset was used to fine-tune the MoulSot.0.3 model for Moroccan Darija speech recognition.

Languages

Moroccan Darija (Arabic script), with natural code-switching to French, English, and Spanish.

Dataset Structure

Data Instances

A typical transcribed instance includes a 16 kHz mono audio file, detailed quality metadata, and its corresponding transcription text. Untranscribed instances include the audio and metadata scores.

Data Splits

• full: The complete 1,500-hour pool of processed Moroccan Darija speech.
• train (100-gt-2.5): High-quality transcribed training subset (~80 hours) with a perceptual quality score (pesq_hyp) > 2.5.
• test: default subset.

Dataset Creation

Curation Rationale

Moroccan Darija remains under-resourced despite being spoken by over 30 million people. MoulSot aims to provide a diverse, high-quality corpus for building robust ASR systems without manual labeling bottlenecks.

Source Data

• Initial pool: ~3,570 videos from 51 YouTube channels.
• Filtering: Filtered based on duration (3-25s), PESQ > 1.0, and Audiobox Aesthetics metrics (Production Quality > 5.0, Content Usefulness > 4.0).
• Diarization: Pyannote 3.1 was used to prioritize single-speaker segments.

Annotations

Transcriptions were generated automatically using Google's Gemini 2.5 Pro, tested to be the most consistent at providing accurate transcriptions in Arabic script while preserving original Latin script for French/English loanwords.

Citation Information

If you use MoulSot-Full in your research, please cite:

@misc{moulsot2026,
  title   = {MoulSot: A Curated Moroccan Darija Speech Dataset and Fine-Tuned ASR Model},
  author  = {Atlasia},
  year    = {2026},
  url     = {https://huggingface.co/atlasia/moulsot.v0.3}
}