Improving speech recognition robustness using non-standard windows

Abstract

The windowing problem of short-time frequency analysis in speech recognition systems (SRSs) is considered. Design possibilities for different non-standard window sequences are presented. The traditional "digital filtering" approach to the design of finite window sequences with linear and nonlinear phase response is examined. Since human hearing is relatively insensitive to phase distortions of speech signals, ideas of alternative windows with nonlinear phase response are also investigated. The two most promising design methods for nonlinear phase windows are discussed. Practical performance comparison of such windows with the Hamming window on two real SRSs is presented. They show that the non-standard window sequences can contribute to greater SRS robustness. Additional research on non-standard windows and the parameterization process as a whole is suggested.