moulsot.v0.3

MoulSot v0.3 — Automatic Speech Recognition for Moroccan Darija

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/65f5c3528fb2b1535728138f/ZcWPJ5cxaX-bh14-0wTVG.png" width="300" /> </div>

Model Description

MoulSot v0.3 is an Automatic Speech Recognition (ASR) model developed by Atlasia for Moroccan Darija, designed to handle real-world speech with strong robustness to code-switching (Darija ↔ French ↔ Arabic).

The model is trained on a large, curated dataset and aims to provide high-quality transcription for conversational, media, and user-generated audio in Moroccan dialect.

• Model type: End-to-end ASR
• Language(s): Moroccan Darija (primary), with code-switching (French, English, Arabic)
• Version: v0.3
• Repository: https://huggingface.co/atlasia/moulsot.v0.3
• Data pipeline: https://github.com/atlasia-ma/MoulSot

---

Key Features

• 🔊 1,500 hours of curated Darija audio
• 🏆 80 hours of gold-standard transcriptions
• ⚡ Native support for code-switching
• 🎙️ Designed for real-world noisy and conversational speech

---

Training Data

Overview

The MoulSot dataset is built from a large-scale data pipeline focused on collecting, filtering, and curating Moroccan Darija audio.

• Total audio: ~1,500 hours

• High-quality labeled subset: 80 hours

• Speech characteristics:

  • Spontaneous and conversational
  • Multi-domain (media, informal speech, etc.)
  • Code-switching between Darija, English, French, and Modern Standard Arabic

Data Pipeline

The full data processing pipeline is open-sourced: 👉 https://github.com/atlasia-ma/MoulSot

It includes:

• Audio collection and aggregation
• Cleaning and normalization
• Segmentation
• Annotation workflows
• Quality filtering

---

Training Procedure

While full technical details are released progressively, the model training includes:

• Supervised training on high-quality labeled data (80h)
• Optimization for multilingual/code-switched contexts
• Iterative refinement across versions

More details will be available in the technical blog.

---

Intended Use

Primary Use Cases

• Transcription of Moroccan Darija audio
• Voice interfaces and assistants
• Media and content indexing
• Call center and conversational AI
• Speech analytics in Moroccan context

Out-of-Scope Use

• Critical decision-making systems without human validation
• Languages outside Darija/English/French/Arabic code-switching context
• Highly specialized domains without adaptation

---

Performance

MoulSot v0.3 is positioned as a state-of-the-art Darija ASR system, with strong qualitative performance in:

• Conversational fluency
• Handling mixed-language speech
• Robustness to accents and informal usage

Benchmark Comparison

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/65f5c3528fb2b1535728138f/1muOZ4gxuei2ahbQrbWCZ.jpeg" /> </div>

Results (↓ lower is better):

• MoulSot v0.1 — WER: 66.78 / CER: 21.30
• MoulSot v0.3 — WER: 38.99 / CER: 12.58
• ISMA — WER: 39.20 / CER: 13.47

MoulSot v0.3 significantly improves over v0.1 and achieves state-of-the-art performance, slightly outperforming ISMA on both WER and CER.

Leaderboard Ranking

MoulSot v0.3 ranks #1 on the Moroccan Darija ASR leaderboard:

👉 https://huggingface.co/spaces/abdeljalilELmajjodi/moroccan_darija_asr_leaderboard

Key takeaways:

• 🥇 Ranked 1st place among evaluated Darija ASR models
• 📊 Confirms strong real-world performance beyond internal benchmarks
• 🔍 Validates improvements over previous versions and competing systems

Quantitative benchmarks and detailed evaluations will be expanded in the technical report.

---

Limitations

• Performance may degrade on:

  • Extremely noisy audio
  • Rare dialectal variations
  • Domain-specific jargon

• Code-switching handling is strong but not perfect

• Limited publicly available evaluation benchmarks for Darija

---

Ethical Considerations

• The dataset is curated with attention to quality and representativeness, but may still contain biases.

• Users should:

  • Validate outputs in critical applications
  • Be mindful of potential transcription inaccuracies
  • Avoid misuse in surveillance or harmful contexts

---

Usage

Install (Python)

uv pip install qwen_asr

Inference (PyTorch / CUDA)

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "atlasia/moulsot.v0.3",
    dtype="bfloat16",
    device_map="cuda",
)

result = model.transcribe(audio="your_audio.wav", language="Arabic")
print(result.text)

---

MLX (Apple Silicon)

Install:

uv pip install mlx-audio

Run:

from mlx_audio.stt.utils import load

model = load("atlasia/moulsot.v0.3")
audio_path = "your_audio.wav"

transcription = model.generate(audio_path).text
print(transcription)

---

Citation

@misc{moulsot2026,
  title={MoulSot v0.3: The New Champ of Darija ASR},
  author={Atlasia},
  year={2026},
  url={https://huggingface.co/atlasia/moulsot.v0.3}
}

---

Resources & Roadmap

Coming soon:

• Full technical blog
• Dataset release
• Training details
• Future model versions

Stay tuned via: 👉 https://atlasia.ma

---

Acknowledgements

Developed by Atlasia as part of ongoing efforts to advance AI for Moroccan languages and dialects.

License

Apache 2.0 — same as the base model.