These words are too difficult for non-native learners to memorize their transcription. We had them read the data set written only in Chinese characters, not in roman transcription, to see what kinds of mispronunciation they would make in the initial segment of the counter. To detect their mispronunciation, first we have it examined by replaying their speech, and then compared the result with their automatically-computed scores. On the basis of native speaker's scores, we defined an absolute threshold value as -29.00 intuitively, as the native speaker's scores were all located within the score around -28.00. Table 3.3 shows the result of HMM-likelihood scores for the phoneme after numerals on the data set. Totally, eleven scores that are marked off in a bold character were detected as mispronunciation. And then, we investigated the correlation between detected errors and human judgements.
