CatBoost Machine Learning Based Feature Selection for Age and Gender Recognition in Short Speech Utterances

Abstract

Lately, with the rapid growth of various technologies, identifying the information of gender and age give short speech utterances has become a necessity for many applications in daily life like human-robot interaction, targeted marketing, identifying suspects in criminal cases, etc. Despite the comprehensive studies carried out to extract descriptive features, the recognition accuracy is still not satisfactory. In this study, an automatic system is proposed to classify age and gender in short speech utterances without depending on the text. Firstly, two groups of features are extracted from each utterance frame, followed by measuring 10 statistical functionals for each extracted feature dimension. After that, the extracted features dimensions are normalized using the Quantile technique. Then, the CatBoost machine is utilized as an important features detection to select the most discriminatory features for speaker age and gender recognition tasks. For classification purposes, the selected feature dimensions are fed into the Support Vector Machine (SVM). Experiments are conducted on the aGender data-set for measuring the suggested system's performance. The unweighted accuracies (UA) of the proposed system for gender, age, in addition to gender & age is 89.62%, 72.29%, and 71.96%, respectively. The achieved results outperform recent results on the same data-set.