SIGNAL ANALYSIS METHOD WITH NON-GAUSSIAN AUTO-REGRESSIVE MODEL

摘要

A signal analysis method including a non-Gaussian auto-regressive model, wherein an input to the autoregressive model (AR) is modelled as a sequence of symbols (I) from a finite alphabet by a finite state stochastic model (FSSM). Probability density functions (pdf) of an input (X) at each time instant are Gaussian pdfs with the same variance (σ21, σ22) for each symbol and with their means (µ1, µ2) decided by the symbols. In preferred embodiments, the FSSM is a Hidden Markov Model HMM, which, on certain occasions depending on the signal characteristic can be reduced to a Gaussian Mixture Model (GMM). The method preferably includes an identification step of performing an expectation-maximization (EM) algorithm. In the case that the observed signal is noisy, such an EM algorithm may involve an optimal smoothing on the noisy signal with a multi-state minimum mean-square error smoother, such as a soft-decision switching Kalman filter. The method is suited for analysis of both clean and noisy non-Gaussian AR signals that have an impulsive structure in its excitation. Due to its low complexity, it is suitable to apply in equipment with limited signal processing capacity. Especially, the method may form part of an algorithm for applications such as: speech analysis, speech enhancement, blind source separation, blind channel equalization, blind channel estimation, or blind system identification.