Enhancement of noisy speech by maximum likelihood estimation

A noise reduction method for speech restoration is presented. The background noise is assumed to be short-time stationary and the following four cases are considered: (1) the additive noise is white but its variance is unknown; (2) the additive noise is white and its variance is known; (3) the power spectrum distribution is known but its absolute value is unknown; and (4) the power spectrum density is known. The speech signal is modeled as a sum of sinusoidal waves. The basic approach is to consider all of the unknown parameters as random with associated a priori Gaussian probability densities and the estimation procedure is based on the maximum a posteriori approach. Experimental results show that the proposed method is effective