Polyphonic pitch tracking by example

We introduce a novel approach for pitch tracking of multiple sources in mixture signals. Unlike traditional approaches to pitch tracking, which explicitly attempt to detect periodicities, this approach is using a learning framework by making use of previously pitch-tagged recordings as training data to teach spectrum/pitch associations. We show how the mixture case of this task is a nearest subspace search problem which is efficiently solved by transforming it to an overcomplete sparse coding formulation. We demonstrate the use of this algorithm with real mixtures ranging from solo up to a quintet recordings.