図 5.15: Speech rhythm error detection based on templates
As shown in Table 2.3, non-native learners from different countries were chosen to analyze their mora-timed rhythm. They uttered just once a set of data consisting of six words. All speech material was orthographically transcribed before being used for the experiment. They are as follows:
| jikoku (time) | kiNchou (tension) | |
| goudou (action) | kougyou (industry) | |
| kaNkei (relation) | saNsei (agreement) |
Based on the Soft Templates of each word, we investigated whether or not their mora rhythms were correct. The results are illustrated in Figure 5.15. Each mora rhythm is represented in a rectangle and its pattern shows whether the corresponding rhythm is in a normal range: a plain pattern is normal, a gray pattern is too long, and a white and black pattern is too short. Their rhythms looked acceptable on the whole except the words [kiNcho:] and [go:do:]. In particular, learner C and F show a contrastive error for the word [go:do:]. The former produced the second mora in a long duration, whereas the latter did the first mora in a long duration. It will be interesting if we try to measure how well non-native learners from a wide range of L1 backgrounds speak Japanese speech rhythm. Unfortunately, we could not able to investigate this in the experiment because of the lack of non-native learners.
For detected rhythm errors, we asked human judges, comprised of five native speakers, to verify them. By and large, the same tendency was confirmed by human judges. However, for the words that have different meaning according to their accent, the human judges showed different results. For example, [ko:gyo:] has several meanings such as the mining industry and a performance according to its accent. As they considered such an accent in their judgement, in some words we failed to find correlation between automatic error detection and human judgements. Although many statistical studies have shown that unaccented words outnumber accented words in standard (Tokyo) Japanese, we found that the role of accent in Japanese is still important in human communications.
図 5.15: Speech rhythm error detection based on templates
