In this experiment, we investigate the usefulness of automatic scoring provided by the speech recognition techniques to identify whether each phoneme is deviant or not. For this purpose, we prepared three tasks of speech materials. The speech of 8 non-native learners of Japanese was scored by means of automatic scoring in order to calculate quantitative measures of speech quality that are known to be related to perceived fluency. Experiments were carried out to determine whether or not each threshold can be employed to detect critical errors. Experimental tasks and their threshold functions are listed in Table 3.1. Learners involved in the experiments are also listed in Table 2.3. Since, in these experiments, pronunciation quality had to be assessed, five native speakers were selected as human judges. They could listen to the speech fragments as often as they wanted.
ID | DATA SET | SPEECH MATERIAL | THRESHOLD FUNCTION |
Task 1 | M-set | specific Mistakes | Absolute Threshold |
Task 2 | P-set | linguistic disParity | Combined Thresholds |
Task 3 | T-set | prosodic Transfer | Relative Thresholds |