The Baum-Welch training algorithm runs toward local maxima so the likelihood of Ergodic HMM depends on the initial parameters. From such a viewpoint, the relation between the classification rate and the likelihood was examined. In the experiment, the LPC analysis window length was 341.3ms, the code book size was 64 and HMM training was iterated 160 times. Other experimental conditions were the same as in experiment 3. We use the 8 speech data sets and 16 initial models were used for each set. Therefore, 128 experiments were carried out. The results are shown in Fig.6. This figure show that there is a relationship between HMM likelihood and the classification rate.
