부산대학교 도서관

Automatic speaker recognition may be beneficial for children in a wide variety of disciplines, such as child education, security and safety. However, owing to a variety of issues in children’s speech, including immature vocal tracts, native phonology, long pauses, and hesitations, many speech technology applications have trouble recognizing non-native children. In order to assess how the children’s speech impacts the system, the primary objective of this study is to build a closed-set child speaker verffication system for non-native English speakers in both textdependent and text-independent tasks. This study focuses on a Gaussian probabilistic linear discriminant analysis (GPLDA) model that was trained employing i-vectors and several whitening methods. Cosine similarity scoring (CSS) and GPLDA scoring are used to evaluate the claimed outcomes. Additionally, it has been shown that in closedset speaker verification, the zero-phase component analysis (ZCA) transformation outperformed the other whitening transformations.