Voice separation in polyphonic music: A data-driven approach

Much polyphonic music is constructed from several melodic lines - known as voices - woven together. Identifying these constituent voices is useful for musicological analysis and music information retrieval; however, this voice-identification process is time-consuming for humans to carry out. Computational solutions have been proposed which automate voice segregation, but these rely heavily on human musical knowledge being encoded into the system. In this paper, a system is presented which is able to learn how to separate such polyphonic music into its individual parts. This system uses a training corpus of several similar pieces of music, in symbolic format (MIDI). It examines the note pitches in the training examples to make observations about the voice structures. Quantitative evaluation was carried out using 3-fold validation, a standard data mining evaluation method. This system offers a valid solution to this complex problem, with a 12% improvement in performance compared to a baseline algorithm. It achieves an equal standard of performance to heuristic-based systems using simple statistical observations: demonstrating the power of applying data-driven techniques to the voice separation problem.