図 3.6: Learners' average scores and relative thresholds
In Figure 3.6, we look at the results of HMM log-likelihood scores obtained from the utterances of a native model speaker and five non-native learners. The average score was calculated with all the phoneme scores and the occurrence of Figure 3.6 stands for the appearance number of corresponding phoneme in the T-set. Some large score degradation of the learners that is far above the model's score is judged as pronunciation error[19]. On the whole, five kinds of pronunciation were verified by human judges. For the phonemes /ch/ and /z/, most learners have also shown score degradation, they were excluded from human judgement because of the lack of sample number (occurrence is 1). For the long vowels, it will be explained in details in the P-set experiment why they showed score degradation.
図 3.6: Learners' average scores and relative thresholds
