{DIME}: A Continuous Dialect Distance Measure from Classification Embeddings

Dialectal variation in speech is inherently continuous, yet most computational approaches model it as a discrete classification problem. We propose a Dialect Distance Measure derived from embeddings of a dialect classification model (DIME), representing dialectal variation as deviation from a standard reference in embedding space. The score combines distance to a standard reference with projection onto a learned dialect–standard direction. Evaluated on German speech data, the proposed measure shows strong correlation with an independent phonetic reference and remains robust across different reference corpora. In contrast, using the same scoring procedure with off-the-shelf speech embeddings yields less consistent results and does not provide a stable continuous dialectal scale. The suitability of the method for capturing variation is demonstrated using within-speaker as well as between-dialect differences.