A corpus-sensitive algorithm for tonal analysis is described. The algorithm learns a tonal vocabulary and syntax by grouping together chords that share scale degrees and occur in the same contexts and then compiling a transition matrix between these chord groups. When trained on a common-practice corpus, the resulting vocabulary of chord groups approximates traditional diatonic Roman numerals. These parameters are then used to determine the key and vocabulary items used in an unanalyzed piece of music. Such a corpus-based method highlights the properties of common-practice music on which traditional analysis is based, while offering the opportunity for analytical and pedagogical methods more sensitive to the characteristics of individual repertories.