azeddinshr/moroccan-darija-speech-recognition

# Fine-Tuning Whisper for Moroccan Darija Speech Recognition This repository documents experiments with parameter-efficient fine-tuning methods for adapting OpenAI's Whisper model to Moroccan Darija. The work compares LoRA and AdaLoRA approaches against baseline performance. ## Background Moroccan Darija presents unique challenges for automatic speech recognition. Unlike Modern Standard Arabic, which dominates training datasets for Arabic ASR systems, Darija incorporates distinct phonological features, code-switching patterns, and regional variations. This project investigates whether parameter-efficient fine-tuning can effectively adapt pre-trained models to this low-resource dialect. ## Dataset **Source:** Darija Open Dataset v2.0 (Zenodo) The dataset contains spontaneous Moroccan Darija speech across multiple speakers and contexts. **Split configuration:** - Training: 2,000 samples (LoRA-Tiny) / 6,239 samples (AdaLoRA-Small) - Validation: 200 samples (LoRA-Tiny) / 836 samples (AdaLoRA-Small) - Test: 600 samples (evaluation) **Specifications:** - Sample rate: 16kHz - Format: WAV - Annotations: Word-level transcriptions ## Methodology ### Experiment 1: LoRA Fine-Tuning (Whisper-Tiny) **Base model:** openai/whisper-tiny (39M parameters) **Configuration:** ```python lora_config = { "r": 32, "lora_alpha": 64, "target_modules": ["q_proj", "v_proj"], "lora_dropout": 0.05 } ``` **Training parameters:** - Trainable parameters: 589,824 (1.54% of total) - Training samples: 2,000 - Steps: 500 - Learning rate: 1e-3 - Batch size: 8 (gradient accumulation enabled) - Hardware: NVIDIA A100 40GB ### Experiment 2: AdaLoRA Fine-Tuning (Whisper-Small) **Base model:** openai/whisper-small (244M parameters) **Configuration:** ```python adalora_config = { "r": 32, "lora_alpha": 32, "target_modules": ["q_proj", "v_proj"], "target_r": 16, "orth_reg_weight": 0.1 } ``` **Training parameters:** - Trainable parameters: 1,327,968 (0.55% of total) - Training samples: 6,239 - St …