Traditional linguistic analyses, focused on morphological, syntactic and lexical features, as well as phoneme-level differences, divide Finnish into two major dialect groups, that subsequently further split into eight sub-groups. This paper presents a complementary dialectal analysis based solely on acoustic characteristics extracted from an extensive database of spontaneous speech from thousands of speakers from all Finnish dialectal areas. The distances among acoustic characteristics of speech from 17 administrative regions are approximated by prediction accuracies of binary classifiers. The classifiers are trained on principal components extracted from utterance embeddings obtained through a large-scale pretrained neural model. The clustering of regional varieties based on these distances yields geographically meaningful dialectal groupings, largely corresponding to the results of the traditional linguistic analyses. Our subsequent analysis indicates that the clustering makes use of prosodic characteristics of utterances.