5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

On Robust Speech Analysis Based On Time-Varying Complex AR Model

Keiichi Funaki, Yoshikazu Miyanaga, Koji Tochinai

Hokkaido University, Japan

We have already developed time-varying complex AR (TV-CAR) parameter estimation based on minimizing mean square error (MMSE) for analytic speech signal. Although the MMSE approach is commonly and successfully applied in various parameter estimation such as conventional LPC, it is well-known that an MMSE method easily suffers from biased and inaccurate spectrum estimation due to non-Gaussian nature of glottal excitation for voiced speech in the context of speech analysis. This paper offers robust parameter estimation algorithm for the TV-CAR model by applying Huber's robust M-estimation approach and two kinds of robust algorithms are derived: Newton-type algorithm and weighted least squares (WLS) algorithm. The preliminary experiments with synthetic signal generated by glottal source model excitation and natural speech uttered by female speaker demonstrate that the time-varying complex AR method is sufficiently robust against non-Gaussian nature of glottal source excitation owing to the improved resolution in the frequency domain.

Full Paper

Bibliographic reference.  Funaki, Keiichi / Miyanaga, Yoshikazu / Tochinai, Koji (1998): "On robust speech analysis based on time-varying complex AR model", In ICSLP-1998, paper 1001.