A note event-based decomposition of polyphonic recordings applied to single-channel audio source separation

Abstract

In this work, a novel decomposition method for single-channel polyphonic recordings is presented and applied to the separation of harmonic or nearly-harmonic sources from within an audio mixture. The proposed method is based on the iterative estimation and extraction of note events, where each event is defined as a section of audio containing largely harmonic energy identified as coming from a single sound source. In every iteration, the pitch trajectory of the predominant note event is automatically selected from an array of fundamental frequency estimates and used to guide the separation of the event’s spectral content, using two different methods: time-frequency masking and time-domain subtraction. A residual signal is then generated and used as the input mixture for the next iteration. This iterative stage continues until a predefined maximum number of iterations is reached. After convergence, all detected note events are clustered into individual sources by exploiting end-user interaction. Performance evaluation is carried out on a set of test recordings, containing excerpts of real music performances with different levels of polyphony, and the obtained results are compared with those generated with an alternative semi-automatic algorithm.