How nassimaouldouali/asr-arabic-finetuning works
This repository is a reproducible Arabic speech-recognition research pipeline: it downloads Common Voice 18, normalizes and filters the audio/transcripts, measures zero-shot baselines, fine-tunes Whisper-small with LoRA, then evaluates and reports the adapted model. The README also describes figures, checkpoint publication, and experiment tracking. The script excerpts show the principal data and model flows; figure-generation and publication details are only partly documented, so uncertain links are omitted or marked unverified.
Main entry points: 04_train.py LoRA fine-tuning · best Best LoRA checkpoint · 01_download.py Dataset download · 03_evaluate_baselines.py Zero-shot baselines
Components 16
Data preparation
- Dataset download
src/01_download.py - Raw dataset
- Arabic preprocessing
src/02_preprocess.py - Processed splits
Modeling
- LoRA fine-tuning
src/04_train.py - Best LoRA checkpoint
src/checkpoints/best
Evaluation
- Zero-shot baselines
src/03_evaluate_baselines.py - Full-set evaluation
src/05_evaluate_finetuned.py
Research outputs
- Evaluation reports
results - Research figures
src/06_generate_figures.py - Generated figures
src/images
Other components
- Researcher
- Common Voice 18
- Whisper-small
- Weights & Biases
- Hugging Face Hub
Connections 19
- Researcher → Dataset download: runs pipeline
README.md - Common Voice 18 → Dataset download: provides splits
src/01_download.py - Dataset download → Raw dataset: saves dataset
src/01_download.py - Raw dataset → Arabic preprocessing: loads dataset
src/02_preprocess.py - Arabic preprocessing → Processed splits: saves splits
src/02_preprocess.py - Processed splits → Zero-shot baselines: supplies test set
src/03_evaluate_baselines.py - Whisper-small → Zero-shot baselines: provides model
src/03_evaluate_baselines.py - Zero-shot baselines → Evaluation reports: writes metrics
src/03_evaluate_baselines.py - Processed splits → LoRA fine-tuning: supplies training data
src/04_train.py - Whisper-small → LoRA fine-tuning: provides base weights
README.md - LoRA fine-tuning → Best LoRA checkpoint: saves adapter
- LoRA fine-tuning → Weights & Biases: logs experiments
README.md - Processed splits → Full-set evaluation: supplies test set
src/05_evaluate_finetuned.py - Best LoRA checkpoint → Full-set evaluation: loads adapter
src/05_evaluate_finetuned.py - Full-set evaluation → Evaluation reports: writes reports
src/05_evaluate_finetuned.py - Processed splits → Research figures: reads distributions
src/06_generate_figures.py - Raw dataset → Research figures: reads raw text
src/06_generate_figures.py - Research figures → Generated figures: saves plots
src/06_generate_figures.py - Best LoRA checkpoint → Hugging Face Hub: published model
README.md