GLoRIA: Gated Low-Rank Interpretable Adaptation for Dialectal ASR

A research team led by Pouya Mehralian has introduced GLoRIA, a novel parameter-efficient adaptation framework designed to tackle the persistent challenge of automatic speech recognition in dialect-heavy environments. The system addresses the dual problems of strong regional variation and limited labeled data by leveraging metadata—specifically geographic coordinates—to intelligently modulate low-rank updates within a pre-trained encoder. On the GCND corpus, GLoRIA outperformed existing methods including geo-conditioned full fine-tuning, standard LoRA (Low-Rank Adaptation), and both dialect-specific and unified full fine-tuning approaches, establishing new state-of-the-art word error rates. The work, accepted for presentation at ICASSP 2026, demonstrates that metadata-guided adaptation can be more effective than brute-force training.

Technically, GLoRIA injects trainable low-rank matrices into each feed-forward layer of a model. A gating multi-layer perceptron (MLP) determines the non-negative contribution of each LoRA rank-1 component based solely on the input location metadata. This architecture allows the model to specialize for different dialects efficiently, updating fewer than 10% of total parameters. Crucially, GLoRIA shows strong generalization capabilities, performing well on completely unseen dialects and even in extrapolation scenarios. Beyond performance, the framework offers interpretability: the adaptation patterns learned by the gating network can be visualized geospatially, providing insights into how dialect features correlate with geography. This positions GLoRIA not just as a performance tool, but as a research instrument for linguistic study.