Children Age and Gender Classification Based on Speech Using ConvNets

Abstract

In this paper, we present a study about building age and gender automatic classifiers for children at their first school years (between 6 and 11 years old).We created a speech corpus with 174 children interacting with a couple of robots in a Wizard of Oz scenario.The recorded speech was manually segmented and then characterized with low-level acoustic features.Next, we trained the classification models using a convolutional neural network architecture.Due to the complexity in the tuning process for the correct selection of the parameters used for this type of neural network, we integrated the use of a mathematical object called covering arrays to generate the set of optimal parameters for neural network architecture.Given the complexity of the classification of children speech, we obtained encouraging results.Our results indicate that it is difficult to achieve an accurate classification of children with very close ages.By grouping the subjects into two or three ages, the results improved significantly.On the other hand, the task of gender identification was less challenging, and we obtained higher classification performance measures.