Probing Geographic Performance Gaps in Moroccan ASR with ALAMA: Annotated Local Audio of Moroccan Arabic

Moroccan Arabic (Darija) is routinely treated as a single dialect in NLP research, yet it exhibits substantial regional variation across its geography. This pilot study questions the monolithic framing by collecting a regionally annotated speech corpus, 60 minutes of broadcast audio from Casablanca, Tangier, and Oujda, transcribed and verified by native speakers, and evaluating 17 state-of-the-art ASR systems against it. Our results show architecture-dependent performance gaps across regions: systems exhibit significant gaps in WER points between the best- and worst-served regions, and fine-tuned "Moroccan Arabic" checkpoints exhibit inconsistent regional profiles that are hard to reconcile with balanced training data. These findings suggest that national-level dialect labels obscure regionally uneven model coverage with real performance consequences. We release the dataset and support the adoption of more fine-grained provenance labeling as standard metadata for Arabic dialect resources.