Towards Robust Identification of Age-Sensitive Acoustic Features for Developmental Speech Analysis in Low Resource Bodo Language

Abstract

Children’s speech exhibits wide variation because their vocal tract and articulatory system mature progressively with age, and thus, analysis of children's speech faces greater difficulty when compared to adults' speech. When the language is a low-resource language where only scarce child speech corpora are available and the development of speech has not been extensively explored, it becomes extremely difficult to build any sophisticated child speech technologies. This study investigates age-sensitive acoustic characteristics of child’s speech in the low-resource Bodo language using Mel-Frequency Cepstral Coefficients (MFCC) and their dynamic characteristics. The children's speech recordings from various ages groups are analyzed, and the MFCC, mean, MFCC variance, del-MFCC mean, and del-MFCC variance are computed from each of the segments to find age-varying features. Correlation analysis, linear regression and one-way ANOVA were employed to identify the acoustically varying features as per age. From the experiments, it was found that deltamean3, deltamean2, mean3, mean6, and mean13 possess a high association with speech speaker age at all the statistical analyses tests. The developmental profile of selected acoustic features shows that both spectro-temporal characteristics of speech are growing into stability as children progress in their age to demonstrate maturity in the mechanism of speech production. The analysis and research using these age-sensitive acoustic features may provide valuable directions to developmental speech research and help to develop useful child-based speech applications for low-resource languages like Bodo.