AI maps Arabic dialects as continuous space, predicts speaker origin with 481km median error

A team of researchers (Khadraoui, Ammar, Benjdira, Khan, Turki, Boulila) has published a paper on arXiv proposing a regression-based approach to Arabic dialect geolocation. Instead of treating dialects as discrete categories, the model predicts speaker origin as continuous geographic coordinates using a hierarchical neural architecture. It fuses frame-level representations from two pre-trained speech encoders—XLS-R-300M and Whisper-large-v3—with phonotactic descriptors, passing them through a Transformer encoder with learnable attention-pooled query. A spherical geodesic loss optimizes great-circle distance directly on Earth’s surface, avoiding planar distortions.

Under a strict leakage-free 5-fold GroupKFold protocol (grouped by source recording), the model achieves a pooled median localization error of 481.2 km and auxiliary country/city accuracies of 64.5% and 45.2%. A permutation Mantel test on the latent space provides quantitative evidence for the Arabic dialect continuum hypothesis. To test real-world generalization, the team introduces a city-masking protocol where two cities per fold are held out from training but included in validation; under this zero-shot regime, mean error rises to 1173.3 km (a 1.32x degradation). The findings establish continuous geographic modeling as a principled framework for Arabic dialect geolocation while quantifying substantial headroom for improvement.