Rule-based method for pitch level classification for a Japanese pitch accent CALL system

要旨

In order to make a CALL system to detect lexical pitch accent errors, we are developing a multi-level pitch pattern recognition algorithm. With this method, it is possible to detect a wider range of pitch accent errors than with previous methods and it is possible to have more flexibility in displaying the pitch pattern to the learner. This rule-based method is based on perceptual experiments in which we made use of speech synthesis to manipulate the pitch of utterances to perform quantitative analysis on accent perception. The recognition is based on two-mora units and the local pitch level pattern for those units. By combining the local pitch level patterns for a word it is possible to derive the two level pitch pattern for the word and also provide a variety of ways to display the word to the language learner. The average agreement rate with the labeling for these units was 75% in the case of the Japanese database and 66% in the case of the non-Japanese database.

主題