Abstract: Many musics across the world are structured around multiple modes, which hold a middle ground between scales and melodies. We study whether we can classify mode in a corpus of 20,865 medieval plainchant melodies from the Cantus database. We revisit the traditional ‘textbook’ classification approach (using the final, the range and initial note) as well as the only prior computational study we are aware of, which uses pitch profiles. Both approaches work well, but largely reduce modes to scales and ignore their melodic character. Our main contribution is a model that reaches 93–95% F1 score on mode classification, compared to 86–90% using traditional pitch-based musicological methods. Importantly, it reaches 81–83% even when we discard all absolute pitch information and reduce a melody to its contour. The model uses tf–idf vectors and strongly depends on the choice of units: i.e., how the melody is segmented. If we borrow the syllable or word structure from the lyrics, the model outperforms all of our baselines. This suggests that, like language, music is made up of ‘natural’ units, in our case between the level of notes and complete phrases, a finding that may well be useful in other musics.