Investigations on the applications of dynamical instabilities and deterministic chaos for speech signal processing

Abstract

This study investigates the potential use of Reconstructed Phase Space (RPS) based parameters for Speech Signal Processing by utilizing nonlinear dynamical systems theory. In this approach features are extracted from the time domain. Study of nonlinear dynamical system shows that the RPS is able to capture the nonlinear information of the underlying system, that cannot be captured by frequency domain analysis. A multimedia based system is converted into a low cost data acquisition system by adding a presampling antialiasing analog filter prior to A/D converter of the sound card. A speech database of short vowels in Malayalam is created. Nonlinear invariant parameters of vowel sounds are calculated. With these parameters one can quantify the chaotic behaviour of the speech signal. Reconstructed Phase Space is generated for speech sounds by the method of time delay embedding. From the reconstructed space, a unique parameter called Reconstructed Phase Space Distribution Parameter (RPSDP) is extracted. These parameters are found to be similar for same vowel and differ from vowel to vowel. They are further used in the recognition experiments. A new method for pitch estimation using Reconstructed Phase Space in two dimensions is presented. The proposed new method does not suffer from the limitations of other short term pitch estimation techniques. The problems in choosing the optimal time delay and the minimum embedding dimension for the reconstruction of phase space using the method of delays are addressed in this thesis. A simple procedure that quantifies expansion from the identity line of embedding space is developed for choosing proper time delay.