A method based on local spectral features and missing feature techniquesudis proposed for the recognition of harmonic sounds in mixtureudsignals. A mask estimation algorithm is proposed for identifyingudspectral regions that contain reliable information for each soundudsource and then bounded marginalization is employed to treat theudfeature vector elements that are determined as unreliable. The proposedudmethod is tested on musical instrument sounds due to theudextensive availability of data but it can be applied on other soundsud(i.e. animal sounds, environmental sounds), whenever these are harmonic.udIn simulations the proposed method clearly outperformed audbaseline method for mixture signals.
展开▼
