Noise-robust speech analysis using system identification methods

This paper proposes a modified linear prediction method for speech analysis, using two system identification methods-the least-square method and the instrument variable method-for the estimation of the coefficients of an all-pole filter. Whereas the linear prediction method estimates the coefficients of all-pole filters from speech signals, which are observed output signals, the system identification method estimates coefficients of all-pole filters from observed output signals and the input signals. This paper derives a novel technique that estimates input signals from speech signals that are observed output signals with a high degree of accuracy and robustness with respect to added noise, by generating improved prediction error signals. The paper also shows that when voiced speech is to be analyzed, if input signals, which are an impulse chain, can be accurately estimated, the estimation of filter coefficients can yield a high degree of accuracy provided that the least-square method is used, and that in this manner, the pitch period dependency can be removed. We also show that by applying the instrument variable method using an auxiliary model, the accuracy of estimation of filter coefficients in a noisy environment can be substantially improved while maintaining the properties of the least-square method. The effectiveness of these system identification methods for speech analysis is demonstrated through computer simulations. (C) 2002 Wiley Periodicals, Inc.