Audio-Based Classification and Geographic Regression of Austrian Dialects

Dialect classification remains challenging due to regional variability and limited dialect-specific datasets. This study addresses these challenges by leveraging a novel dataset of 304 speakers from 108 locations across Austria for automatic classification of Austrian dialects. To minimize speaker-specific biases and enhance dialectal features, speaker augmentation techniques are applied. Classification is conducted at three levels: location, dialect group, and federal state. Additionally, a regression task predicts the speakers' geographic coordinates, with the wav2vec 2.0 model architecture achieving an average test-set distance error of 66.7 kilometers. This work represents a unique approach to fine-grained dialect classification and geographic location prediction for Austria. Finally, model explainability is explored using Integrated Gradients (IG), identifying the most relevant speech segments for classification within each dialect group.