Statistical learning of harmonic movement

We explore the application of statistical techniques, borrowed from natural language processing, to music. A probabilistic method is used to capture a nd generalise from the local harmonic movement of a corpus of seventeenth-c entury dance music. The probabilistic grammars so generated are then used f or experiments in generation (composition). The corpus is preprocessed in a novel way, automatically converting the har monies into a normal form to capture the underlying harmonic similarities b etween pieces. It is then automatically marked up with constituent boundari es (beginnings and ends of pieces, phrases and bars), to enable the learnin g process to capture some of the higher-level structure of the music. The experiment is promising, and a sample of the results are given. We disc uss the limitations of the approach, and how they might be overcome.