This paper addresses the problem of automatic speech recognition in the presence of interfering noise. The approach adopted is to compensate the parameters of a clean speech model given the statistics of the interfering noise. In this work these statistics are assumed to be modelled with a Hidden Markov Model. The basic theory of static coefficient Parallel Model Combination (PMC) is reviewed and placed within the framework of approximating the Maximum Likelihood (ML) estimate of the corrupted speech model, given the clean speech and interfering noise models. A new form of PMC is described which improves the performance of fixed grand variance based recognition schemes, by compensating the variance to be more representative of the corrupted speech fixed grand variance. In addition, the paper examines the problem of compensating delta coefficients in a PMC framework. Expressions for ML estimates of delta coefficients are derived and computationally efficient approximations of these estimates are given. The effectiveness of compensating delta parameters is discussed.

Keywords: speech recognition, noise compensation, HMM, PMC.