Abstract: In this study propose a new approach in using Hidden Markov Models (HMMs) for speech recognition. Although HMMs are the state-of-the art speech recognition systems, they suffer from some inherent limitations. One of these limitations is the independence assumption in the HMMs formalism. In the approach described in this study, we use in the vector quantization process, grouped vectors of different length to explicitly model the natural correlation between adjacent frames, instead of using a single vector in the standard method. The system is tested on an Arabic isolated digits (0-9) recognition task, our method achieves a 21% reduction in word error rate evaluation compared with the standard approach.