BADR-JOULAlI/Darija-ASR

Reproducible ASR research project for Moroccan Darija, focusing on CTC fine-tuning, Arabic/Latin code-switching, tokenizer coverage, and WER/CER evaluation. # Darija ASR: Reproducible Speech Recognition Research Public research overview for automatic speech recognition (ASR) in Moroccan Darija, with particular attention to Arabic script, Latin script, and Arabic-Latin code-switching. > This repository currently contains public documentation and release > guidance only. Private datasets, model weights, checkpoints, infrastructure > details, and non-reviewed experimental outputs are intentionally excluded. ## Abstract Moroccan Darija presents practical challenges for speech recognition because it is represented through multiple writing conventions, includes borrowed Latin vocabulary, and frequently switches scripts within the same utterance. This project develops a reproducible evaluation and adaptation protocol for pretrained speech recognition models, with emphasis on connectionist temporal classification (CTC) systems. The central objective is not only to reduce recognition error, but also to measure where a model succeeds or fails: Arabic-only speech, Latin-only speech, and mixed-script speech must be reported separately rather than hidden inside one global score. ## Research Questions 1. How effectively do pretrained ASR encoders transfer to Moroccan Darija? 2. How does tokenizer coverage affect Arabic, Latin, and mixed-script output? 3. Does source-level data splitting provide a more reliable estimate of generalization than random segment splitting? 4. Which checkpoint provides the best validation WER and CER trade-off? 5. Which error patterns remain after adaptation, especially in code-switching and borrowed Latin vocabulary? ## Experimental Principles Each experiment should follow these principles: - use a fixed random seed and record all configuration values; - separate train, validation, and test data by speaker, recording, or source; - fit or select preprocessing decisions using training data only; - keep the final test set untouched during model and checkpoint selection; - compare models on exactly t …