Recurrence Analysis Of Dynamical Characteristics For Speech Signals

In recent years, with the development of the research in nonlinear dynamical characteristics of speech signals and the nonlinear dynamical theory, nonlinear speech signal processing has become a hot topic. However, theoretically the nonlinear invariants, such as the correlation fractal dimension and the maximum Lyapunov exponent, require for large amounts of data and the pre-assumption of stationarity of the time series, which limit their applications and may result in wrong explanations. Therefore, developing a simpler and more effective method for dealing with the nonstationarity in speech signals and extracting some new features would be the focus of this dissertation.Firstly, this dissertation carries out quantitative and qualitative analysis of speech signals based on the nonlinear dynamical theory. In modeling the speech production, three different physical mechanisms are introduced, e.g. the vocal fold oscillation for voiced speech, the turbulent sound source for unvoiced speech and the interaction phenomena for human whistling. The problem of phase space reconstruction for speech signals is also studied. The false nearest neighbor technique and the average mutual information method are used to calculate the embedding dimension and the delay time mτrespectively, and then the statistical results for a set of speech phones help to choose m andτfor the continuous speech signals. Furthermore, the correlation fractal dimension and the maximum Lyapunov exponent of speech are investigated. These chaotic measures only show the existence of chaos in speech but not an exact quantitative description. Secondly, we introduce the recurrence plot (RP) and the recurrence quantification analysis (RQA) techniques through reviewing the recent techniques for the short nonstationary time series analysis. Both large-scale and small-scale patterns of RPs are investigated. RQA parameters, such as the percentage of recurrence (RR), the percentage of points forming upwards-diagonal lines to recurrent points (DET), the length of the maximum diagonal line (Lmax), are described. In addition, the applications of RP and RQA in nonlinear time series analysis, the correlation analysis between two time series...