Figure 10

Phone accuracy change due to template-matching-based rescoring with respect to HMM baseline. For the vowel and semivowel classes, the KL- and LLR-based template matching obtained better performance than HMM baseline. For the stop class, template matching using the LLR distance got better phone accuracy than HMM baseline while using the KL distance did not help. For the fricative class, the phone accuracies were worse than HMM baseline for both KL- and LLR-based template matching. For the nasal class and silence, the changes between template-based methods and HMM baseline were not significant.