Nonlinear Modeling of Speech Signals Using Models for Chaotic Systems
During my M.Eng. thesis and my first graduate years I explored the use of nonlinear function approximation/machine learning techniques, applied to capturing and analysing the nonlinear dynamics of speech signals.
As in chaotic time series analysis, an embedding of the observed scalar speech time series is generated using a feature vector formed from delayed samples of the signal. For a sufficiently high embedding dimension, geometrical properties of the attractor of the original dynamical system remain intact by this procedure, allowing for the estimation of invariants of its dynamics, like Lyapunov exponents.
My research focused on algorithms for learning the embedded signal dynamics, using a sparse set of input-output pairs. Using mixture-of-experts for regression we obtained good results both on synthetic and real data, and the estimated Lyapunov exponents turned out to be useful features for speech recognition.
Related Publications:
I. Kokkinos and P. Maragos,
Nonlinear Speech Analysis Using Models for Chaotic Systems,
IEEE Trans. on Speech and Audio Processing, Vol. 13(6), pp. 1098-1109, 2005.
V. Pitsikalis, I. Kokkinos and P. Maragos,
Nonlinear Analysis of Speech Signals: Generalized Dimensions and Lyapunov Exponents.,
Proc. European Conference on Speech Communication and Technology (EUROSPEECH), 2003.
P. Maragos, A. Dimakis and I. Kokkinos.
Some Advances in Nonlinear Speech Modelling Using Modulations Fractals and Chaos.
In Proc. IEEE Int'l. Conf. on Digital Signal Processing, 2002.
I. Kokkinos.
Nonlinear Speech Processing Using Models for Chaotic Systems (In Greek), a more extensive report
School of Electrical and Computer Engineering, National Technical University of Athens, 2001