@techreport{oai:ipsj.ixsq.nii.ac.jp:00234654, author = {恩田, 健太郎 and 朴, 浚鎔 and 齋藤, 大輔 and 峯松, 信明}, issue = {42}, month = {Jun}, note = {ある言語の母語話者音声コーパスのみを用いた自己教師あり学習により,その言語に固有の音韻表象と解釈できる離散シンボルを生成し,音声→離散シンボル列→音声と再合成を行うことで,その言語特有の訛りが他言語入力に対して付与されるプロセスを模擬する.出力音声は,特に分節的な観点において,実際の L2 話者の発話と類似した訛りの傾向を持つことが示された., In this paper, we simulate the process of foreign accentuation by generating discrete symbols that can be interpreted as language-specific phonological representations using self-supervised learning only with native speech corpora. The output speech, which is synthesized by the process of speech → discrete symbol sequence → speech, was shown to have similar behaviors of accentuation to that of actual L2 speakers, especially from a segmental perspective.}, title = {母語話者音声コーパスのみを用いた自己教師あり学習による外国語訛り音声の合成}, year = {2024} }