Search papers, labs, and topics across Lattice.
The paper introduces GLoRIA, a parameter-efficient adaptation framework for dialectal ASR that uses location metadata to modulate low-rank updates in a pre-trained encoder. GLoRIA injects low-rank matrices into feed-forward layers and uses a gating MLP conditioned on location to determine the contribution of each rank-1 component. Experiments on the GCND corpus demonstrate that GLoRIA outperforms various fine-tuning and LoRA baselines, achieving state-of-the-art WER with fewer updated parameters and exhibiting strong generalization to unseen dialects.
Location-aware gating of low-rank adaptations unlocks SOTA dialectal ASR with <10% parameter updates and interpretable geospatial patterns.
Automatic Speech Recognition (ASR) in dialect-heavy settings remains challenging due to strong regional variation and limited labeled data. We propose GLoRIA, a parameter-efficient adaptation framework that leverages metadata (e.g., coordinates) to modulate low-rank updates in a pre-trained encoder. GLoRIA injects low-rank matrices into each feed-forward layer, with a gating MLP determining the non-negative contribution of each LoRA rank-1 component based on location metadata. On the GCND corpus, GLoRIA outperforms geo-conditioned full fine-tuning, LoRA, and both dialect-specific and unified full fine-tuning, achieving state-of-the-art word error rates while updating under 10% of parameters. GLoRIA also generalizes well to unseen dialects, including in extrapolation scenarios, and enables interpretable adaptation patterns that can be visualized geospatially. These results show metadata-gated low-rank adaptation is an effective, interpretable, and efficient solution for dialectal ASR.
