Abstract
A novel approach using non-negative matrix factorization (NMF) for onset detection of musical notes from audio signals is presented. Unlike most commonly used conventional approaches, the proposed method exploits a new detection function constructed from the linear temporal bases that are obtained from a non-negative matrix decomposition of musical spectra. Both first-order difference and psychoacoustically motivated relative difference functions of the temporal profile are considered. As the approach works directly on input data, no prior knowledge or statistical information is thereby required. A practical issue of the choice of the factorization rank is also examined experimentally. Numerical examples are provided to show the performance of the proposed method. © 2006 IEEE.