Over the past few years, there have been several published reports on the use of Maximum Mutual Information (MMIE) for training HMM parameters. Lately, some reports have appeared, proposing different solutions to avoid the computational cost associated with the low convergence of the optimization technique. This paper proposes a new method for increasing the velocity of convergence, by pre-adapting the Codebook Weights of multiple codebooks in Semi-Continuous HMM (SCHMM). Experimental results on a small vocabulary recognition task show not only a fast convergence but also a 21% error rate reduction.