Reduction of Musical Noise in Spectral Subtracted Speech Using Excitation Source Information

Abstract

During speech production, the significant excitation of the vocal tract occurs at places like instants of glottal closure and the onset of events like bursts. Hence the energy in the excitation signal is high around these places and these instants play an important role in the perception of speech. This property is exploited in this work to derive the weight function. The linear prediction (LP) residual of the spectral subtracted speech signal is weighted by the weight function to enhance the significant excitation of speech. The musical noise in the enhanced speech obtained by exciting the time varying all pole filter with the modified residual is found to be reduced significantly