Speech enhancement based on the response features of facilitated EI neurons

Abstract

A real-time approach for the enhancement of speech at zero degree azimuth is proposed. This is achieved inspired by the response features of the "Facilitated EI neurons". This way, frequency segregation through a bandpass filter bank is followed by "supression analysis" which inhibits sources that are not at "facilitated" positions. Unlike with the existing approaches for the solution of cocktail party problem, where the performance under low SNR (signal-to-noise ratio) reverberation conditions is severely limited, the proposed approach has the capability to circumvent these problems, This is quantified through both objective and subjective performance measures and supported by real world simulation examples. © Springer-Verlag Berlin Heidelberg 2006.