Role of Acoustics and Prosodic Features for Children's Age Classification

A speech signal contains numerous details about the speaker, including their gender, age, accent, and health. In recent times, deep neural networks have shown remarkable abilities in identifying the gender and age of human speech signals. Still, not much research has been done on using speech signals to determine a child's age. In this paper, we have explored the role of acoustics and prosodic features in children's age classification. Further, 23 unique combinations of acoustics and prosody features were also explored to boost the classifier's performance. The databases utilized for the age classification experiments are the PF-Star and CMU Kids children's speech corpus. For the PF-Star dataset, the feature combination of MFCC and Tempo, demonstrated the best classification accuracy of 88.64%. Whereas our best feature combination, containing MFCC, Chroma and Spectral features, achieved the best classification accuracy of 94.73% using an ANN classifier. The aforementioned unique feature combinations show an absolute improvement of 4.55% and 1.46%, in classification of children's age for PF-Star and CMU Kids respectively, over the baseline performances.