The aim of the work described in this paper is to make a state of the art HMM based speech recognition system robust to the effects of additive noise. It describes the application of Parallel Model Combination to a large vocabulary speech recognition in noise task, the ARPA 1994 CSRNAB Spoke 10. Two variants of PMC have been examined; the Log-Add approximation and Data-driven Parallel Model Combination (DPMC). In the latter two schemes were investigated, one compensating both the means and variances, the other just the means. All the schemes were found to improve the noise robustness of the system. Comparable performance to the DPMC mean compensated system was achieved using the Log-Add approximation. However, little gain in performance was found when the variances were compensated, indeed on the evaluation data the performance was degraded. This was surprising, since experiments on the Resource Management database with additive noise showed that variance compensation consistently improved performance.