Non-linear techniques for robust speech recognition

An acoustic speech digital signal can be regarded as a random process repeatedly alternating stationary segments with non-stationary ones. However, the current features used in the mostly recognition system are drawn based linear model theory and are hardly to describe non-stationary character. Consequently, some syllables can not be distinguished in speech parameter space with dimensions as high as 50. This paper tries to develop several features to describe non-stationary measure, a trend degree and the character of continuously exchange between stationary pieces and non-stationary pieces with the help of statistical theory and non-linear random model, a type of doubly random time series model. Our experiments have shown the proposed feature to increase the recognition accurate and the ability of adaptation and self-organisation of the system.