Recently, the vector Taylor series (VTS) approach was proposed as an efficient method for robust speech recognition under various environmental conditions. The VTS approach makes an approximation to the speech contamination procedure by a linearized model and estimates the parameter values using the expectation-maximization (EM) algorithm. In this paper, we apply the VTS approach to environment compensation with assumed noise statistics. In addition, we present a Bayesian adaptation technique with which we can incorporate the a priori knowledge about the noise statistics to the parameter estimation procedure.