A sequential algorithm for robust parameter estimation and enhancement of noisy speech

The sequential estimation of robust parameters and enhancement of noisy speech based on the expectation-maximization (EM) algorithm is considered when the excitation source is a mixture of two Gaussian distributions with different variances. Methods of stochastic approximation are used to convert a traditionally iterative algorithm into a sequential one. The proposed algorithm is robust against the observation noise and enhances noisy speech to produce results that are 3~4-dB better than conventional methods