A Neural Greedy Model for Voice Separation in Symbolic Music

Music is often experienced as a simultaneous progression of multiple streams of notes, or voices. The automatic separation of music into voices is complicated by the fact that music spans a voice-leading continuum ranging from monophonic, to homophonic, to polyphonic, often within the same work. We address this diversity by defining voice separation as the task of partitioning music into streams that exhibit both a high degree of external perceptual separation from the other streams and a high degree of internal perceptual consistency, to the maximum degree that is possible in the given musical input. Equipped with this task definition, we manually annotated a corpus of popular music and used it to train a neural network with one hidden layer that is connected to a diverse set of perceptually informed input features. The trained neural model greedily assigns notes to voices in a left to right traversal of the input chord sequence. When evaluated on the extraction of consecutive within voice note pairs, the model obtains over 91% F-measure, surpassing a strong baseline based on an iterative application of an envelope extraction function.