Mitigation of reverberation on speaker identification via homomorphic filtering of the linear prediction residual

This paper presents a method for mitigating the impact of reverberation upon speaker identification. In particular, two reverberation mitigation techniques were studied: one that applies temporal homomorphic filtering techniques to the Linear Prediction (LP) residual and one that uses Nuisance Attribute Projection (NAP). Typically, linear channel distortion mitigation approaches that use homomorphic filtering assume that the temporal trajectories of the cepstral components have zero-mean for undistorted speech; however, this approximation is rather poor, particularly for any cepstral components that are heavily influenced by the vocal-tract resonances. By processing the LP-residual, where these resonances have been suppressed, a better approximation to the zero-mean assumption is obtained. Using homomorphic filtering of the LP-residual resulted in relative Speaker Identification (SID) improvement of 28%, while applying NAP resulted in up to 27% relative improvement, and further improved the LP-residual processing results by an additional 8%.