In the previous experiment, the model with high likelihood gave the highest speaker classification rate. We tested this relationship in the following experiment. Sixteen different initial models were constructed at random. After Baum-Welch learning, the model with the highest likelihood was selected. Fig.7 shows the average classification rates. The LPC analysis window length is set at 341.3 ms. The universal code book size was varied from 32 and 256. Other experimental conditions were the same as in experiment 3. As can be seen from these results, the average classification rate is 78.8% which means that the performance is improved by about 10%.
