Ask a voice assistant to navigate to a small local restaurant and you'll quickly find where ASR falls apart — long-tail POI names that never showed up in the LM training corpus. This paper from Meituan tackles the problem head-on with a geographical model that biases the recognizer toward POIs near the user's actual location, treating GPS as a first-class contextual signal instead of a post-hoc reranking hint.

The design combines a location-aware language model with a shortlist mechanism that dynamically weights candidate POI names based on regional prior probabilities. On real in-car and mobile queries, the approach cuts POI error rates significantly compared to a strong flat LM baseline, without hurting general-domain accuracy. The walkthrough covers how the geographic priors are estimated, how they're fused into decoding, and the engineering tradeoffs of shipping this in a production nav app. If you build voice interfaces for maps, delivery, or ride-hail, this is a practical playbook you can steal from.