Leveraging Geographic Metadata for Dialect-Aware Speech Recognition

Automatic Speech Recognition (ASR) systems struggle with dialectal variations due to significant phonetic, morphosyntactic, and lexical differences across regions. Traditional approaches rely on dialect-specific models, per-dialect fine-tuning, or unified models that struggle to generalize across diverse dialects. In this work, we propose novel methods for integrating geographical metadata into ASR models to enhance dialect adaptation. By leveraging latitude and longitude as continuous inputs, our model learns smooth transitions across dialectal variations, allowing it to interpolate between known dialects and generalize to unseen regions. Our experiments demonstrate that these approaches significantly outperform both dialect-specific models and state-of-the-art Whisper models while reducing computational and maintenance costs. These findings suggest that location-based adaptation is a promising direction for improving ASR performance in dialect-rich environments.