An Investigation of Feature Difference Between Child and Adult Voices Using Line Spectral Pairs
Ginji Hayashi, Shigeru Katagiri, Xugang Lu, Miho Ohsaki · 2022
The fundamental and formant frequencies of children’s speeches (voices) are significantly different from those of adult speeches due to the difference in vocal tract length between them. Therefore, when a speech pattern recognizer for children’s speech is designed using adult speech, the recognition rates of the children’s speech patterns are significantly reduced. To solve this problem, we focus in this paper on the differences in the Line Spectral Pairs (LSP) frequencies between the speeches of adults and children and analyze them. We use the LSP frequencies by developing a speech recognizer with the Conjugate Structure-Algebraic Code Excited Linear Prediction method, which simultaneously enables both speech recognition and synthesis. In experiments, we analyze the LSP frequencies of the speech data of 10 adults and 10 children, and find that the main differences between children and adults’ speeches are from the frequencies of lower LSP parameters.