We present a methodology for assessing similarities and differences between language varieties and dialects in terms of prosodic characteristics. A multi-speaker, multi-dialect WaveNet network is trained on low sample-rate signal retaining only prosodic characteristics of the original speech. The network is conditioned on labels related to speakers’ region or dialect. The resulting conditioning embeddings are subsequently used as a multi-dimensional characteristics of different language varieties, with results consistent with dialectological studies. The method and results are illustrated on a Swedia 2000 corpus of Swedish dialectal variation.