A deep learning model for gender recognition using voice data

Abstract

Gender recognition using speech signals has become essential due to the advancement in digital technology and the need for computer  systems to be able to classify gender using voice information. Numerous studies have been conducted with an emphasis on enhancing  feature extraction and development of better classifiers for gender recognition based on speech. Out of all the different kinds of models  developed, the LSTM model yields the greatest results. Additionally, for various signal to noise ratios, the LSTM model showed  outstanding generalization performance. However, LSTM models use feed-forward neural networks that has limitations in capturing  frequency and temporal correlations. This paves the way for further research into alternate recurrent-network techniques, which have  been demonstrated to handle contextual information better, in order to achieve additional performance gains. The study improves  gender recognition using a Bi-LSTMLSTM architecture and voice data. The study adopts Relief-based method for feature selection. The  results show that the BiLSTMLSTM model achieved better gender recognition than LSTM-LSTM model at an accuracy of 99.30%, sensitivity  of 99.60% and specificity of 99.00%. The BiLSTM model is successful in achieving higher accuracy and sensitivity values than  LSTM at 1.00% and 2.20% respectively. The model also outperformed classical machine learning approaches (Fine Tree, K Nearest  Neighbor, Linear Discriminant, Logistics regression and Support Vector Machine) in terms of accuracy at a minimum of 2.20% to a maximum of 05%. The comparative analysis of the classification performance shows that deep learning approaches are more successful  in gender recognition than classical machine learning models.