Speech enhancement based on extended Kalman filter and neural predictive hidden Markov model

To represent the nonlinear and nonstationarity nature of speech, we assume that speech is the output of an NPHMM combining a neural network and hidden Markov model (HMM). The NPHMM is a nonlinear autoregressive process whose time-varying parameters are controlled by a hidden Markov chain. Given some speech data for training, the parameter of NPHMM is estimated by a learning algorithm based on the combination of Baum-Welch's algorithm and a neural network learning algorithm using the backpropagation algorithm. A recursive method using a extended Kalman filter with the parameters of a trained NPHMM is developed for enhancing speech signals degraded by statistically independent additive noise assumed to be white Gaussian. Our recursive speech enhancement method achieves an improvement over the method hidden filter model of about 1-1.5 dB in SNR.