mikane2311/asr-darija

# Moroccan Darija ASR ### Fine-Tuning and Deploying an Automatic Speech Recognition System for Moroccan Darija A complete end-to-end ASR project for **Moroccan Darija**, including **dataset preparation**, **model fine-tuning**, **quantitative evaluation**, **inference**, **interactive demo**, and **Dockerized deployment**. --- ## Highlights - Fine-tuned a pretrained **Moroccan Darija ASR** model on the **DODa** dataset - Evaluated performance using **WER** and **CER** - Compared the **base model** and the **fine-tuned model** - Built a local **Gradio demo** for interactive transcription - Packaged the demo using **Docker** - Organized the project with notebooks, reports, and deployment scripts --- ## Project Overview Automatic Speech Recognition (ASR) has significantly improved in recent years thanks to deep learning and pretrained speech models. However, low-resource dialects such as **Moroccan Darija** remain insufficiently supported. This project aims to address that gap by adapting a pretrained ASR model to Moroccan Darija using the **DODa audio dataset** and the transcription target **`darija_Arab_new`**. The project covers the full ASR workflow: - dataset loading and preparation, - fine-tuning of a pretrained model, - evaluation with standard metrics, - qualitative and quantitative comparison, - inference on unseen audio, - demo interface for transcription, - Docker-based packaging. --- ## Objectives The main goals of the project are: - build a functional ASR pipeline for Moroccan Darija, - fine-tune a pretrained Darija ASR checkpoint, - evaluate transcription quality using **WER** and **CER**, - compare performance before and after fine-tuning, - implement an inference pipeline for audio transcription, - develop an interactive demo application, - package the demo in a reproducible way. --- ## Dataset The dataset used in this project is the **DODa audio dataset**, which provides Moroccan Darija speech samples with multiple transcription forms. …