This paper proposes a HMM training algorithm which is based on grouping multiple observations by multiple correlation coefficient . It avoids computing the conditional probabilities directly and considers the correlativity between successive observation vectors and need not increase computation works any more compare...