nassem-asr-darija

MoulSot v0.3 MLX Q4

This repository contains a quantized MLX packaging of MoulSot v0.3, the Moroccan Darija ASR model used by Nassem for local transcription workflows.

Model summary

• Base model: Qwen/Qwen3-ASR-1.7B
• Task: Automatic Speech Recognition
• Languages: Moroccan Darija, Arabic, French, English
• Format: MLX quantized model package for local inference
• License: Apache-2.0

Intended use

This model is intended for local, offline transcription of Moroccan Darija and code-switched speech inside the Nassem workflow.

Notes

• The repo is private.
• The files in this repository are the quantized model artifacts, not a hosted inference endpoint.
• Use this model in the app after it has been downloaded to the device or local runtime storage.

Quick start

Install the MLX ASR helper package and load the model:

uv pip install mlx-audio

from mlx_audio.stt.utils import load

model = load("kamalaitbrahim/nassem-asr-darija")
result = model.generate("your_audio.wav")
print(result.text)

Limitations

• Performance may vary with noisy audio, accents, and extreme code-switching.
• This repository does not provide a cloud inference API.

License

Apache-2.0

Nassem web speech engine preview

The web speech engine packages are available on the web-engine-v1 release branch. This includes platform-specific model manifests and the macOS developer-preview installer.

Nassem runs in the browser and connects to the speech engine installed locally on the laptop. The web preview is kept on its own release branch to preserve compatibility with existing iOS and Android model downloads. See the web setup and validation notes for current test results and release limitations.