3 Binary classification HMM
In this experiment we used the utterances from Thai drama corpus without emotional segmentation to train sixteen-state and two-Gaussian mixtures HMMs for binary classification as mentioned in section III. Then we tested these four classes of HMMs with two data sets from Thai drama corpus and non-professional voice actor. Assumedly, the first test data set is represented for
speaker dependent evaluation and another set is provided for speaker independent evaluation. The results of this binary classification are shown in Table 6. The accuracy of emotional classification can be improved by nearly 2.50% comparing with the baseline system. The accuracy of this classification dropped to 22% when testing in speaker independent.