oussamaaxkos/Asr_darija

# Qwen3-ASR Fine-Tuning for Moroccan Darija Fine-tune Qwen3-ASR on the MoulSot-Full dataset — a curated corpus of Moroccan Darija speech. Based on the MoulSot project, which built a full pipeline crawling ~1,500 hours from YouTube and transcribing 80 hours using Gemini 2.5 Pro. --- ## Project Structure ``` qwen3_asr_darija/ ├── config.py # All paths and hyperparameters ├── check_env.py # Verify GPU and installed packages ├── prepare_data.py # Download dataset and build JSONL manifests ├── train.py # Fine-tune using the official Qwen3-ASR SFT script ├── evaluate.py # Evaluate a checkpoint with WER / CER / RTF ├── infer.py # Transcribe new audio files └── requirements.txt ``` --- ## Requirements - Python 3.10+ - CUDA-capable GPU (16 GB VRAM recommended; T4 works for the default config) - Git Install Python dependencies: ```bash pip install -r requirements.txt ``` --- ## Quick Start ### 1. Check your environment ```bash python check_env.py ``` ### 2. Prepare the data Downloads MoulSot-Full from Hugging Face, resamples audio to 16 kHz, and writes JSONL manifests. ```bash python prepare_data.py ``` By default this uses **1,000 training** and **200 test** samples. Edit `config.py` to change these values. ### 3. Fine-tune the model ```bash python train.py ``` This will: 1. Clone the official `QwenLM/Qwen3-ASR` repository. 2. Apply the fp16 compatibility patch. 3. Launch the SFT training script. Checkpoints are saved to `model_output/`. ### 4. Evaluate ```bash python evaluate.py ``` Optional arguments: ```bash python evaluate.py --checkpoint model_output/checkpoint-250 --samples 100 ``` Reports WER, CER, and mean RTF. Per-sample results are saved to `data/eval_details.tsv`. ### 5. Transcribe new audio ```bash python infer.py path/to/audio.wav python infer.py audio1.wav audio2.wav --checkpoint model_output/checkpoint-250 ``` --- ## Configuration All settings live in `config.py`. Common things to cha …