In this paper we propose a multiple branch hidden Markov model(MBHMM) which is different from the conventional ones. In the basic HMMs, there is only one transition branch from one state to another one. Our new model has multiple transition branches between two states. As a result, it can hold much more spectral information in the speech signal than the basic HMMs. The evaluation, decoding, and training algorithms associated with MBHMM are also derived. The resulting recognizer is tested on a vocabulary of ten Chinese digits over 20 speakers. The recognition results show that MBHMM significantly outperforms the conventional discrete HMM(DHMM).