A novel speech-based model successfully distinguishes between nonfluent, logopenic, and semantic variants of primary progressive aphasia.