Discovering Convolutive Speech Phones Using Sparseness and Non-negativity

Abstract Discovering a representation that allows auditory data to be parsimoniously represented is useful for many machine learning and signal processing tasks. Such a representation can be constructed by Non-negative Matrix Factorisation (NMF). Here, we present a convolutive NMF algorithm that includes a sparseness constraint on the activations and has multiplicative updates. In combination with a spectral magnitude transform of speech, this method discovers speech phones that exhibit sparse activation patterns, which we use in a supervised separation scheme for monophonic mixtures.