Automatic Geminate Insertion Algorithm for Japanese Audio Data

Abstract

Generally, it is quite difficult for Japanese language learners to acquire Japanese special morae, namely, geminate, syllabic nasals and long vowels compared to independent morae. Among these three special morae, geminate is particularly difficult, and it takes much longer to fully acquire both production and perception of it. Especially for learners of Chinese native speakers, previous studies has shown that both production and perception of geminate are difficult in terms of the fact that not only no geminate is found in Chinese language, but also the phonological interaction between Japanese accent and Chinese tones. However, in the field of Japanese speech acquisition, research has not making progress because of a major problem, that is, researchers themselves manually create the acoustic experiment stimuli. Therefore, in this study, as a method to solve this problem, we propose an algorithm that automatically inserts geminate into the audio data used in Japanese speech acquisition research. This algorithm automates the insertion of geminate by performing three processes in order: mora extraction by noise removal, matching of original audio data and extracted mora, and insertion of soundless duration and geminate. The algorithm makes it possible to remove the noise, which is -50 dBFS and continues for 10ms or more, and replace it with soundless duration instead, allowing Japanese native speakers to percept it as geminate. The accuracy was equivalent as a result of comparing the data that was manually modified by a phonology researcher with the data that was generated by the algorithm. The result shows that the algorithm can be a practical solution for the automation of geminate insertion.