Speaker Age and Gender Estimation Based on Deep Learning Bidirectional Long-Short Term Memory (BiLSTM)

Abstract

Estimating the age and gender of the speaker has gained great importance in recent years due to its necessity in various commercial, medical and forensic applications. This work estimates the speakers gender and ages in small range of years where every ten years has been divided into two subcategories for a span of years extending from teens to sixties. A system of speaker age and gender estimation uses Mel Frequency Cepstrum Coefficient (MFCC) as a features extraction method, and Bidirectional Long-Short Term Memory (BiLSTM) as a classification method. Two models of two deep neural networks were building, one for speaker age estimation, and the other for speaker gender estimation. The experimental results show that the deep neural network model of age estimation achieves 94.008 % as accuracy rate, while the deep neural network model of gender estimation achieves 90.816% as accuracy rate.