Novel hybrid DNN approaches for speaker verification in emotional and stressful talking environments

In this work, we conducted an empirical comparative study of the performance text-independent speaker verification in emotional and stressful environments. This work combined deep models with shallow architecture, which resulted novel hybrid classifiers. Four distinct were utilized: neural network-hidden Markov model (DNN-HMM), network-Gaussian mixture (DNN-GMM), Gaussian model-deep network (GMM-DNN), hidden (HMM-DNN). All based on implemented architecture. The used three speech datasets: a private Arabic dataset two public English databases, namely Speech Under Simulated Actual Stress (SUSAS) Ryerson Audio-Visual Database Emotional Song (RAVDESS). test results aforementioned demonstrated that proposed HMM-DNN leveraged Results also showed outperformed all other terms equal error rate (EER) area under curve (AUC) evaluation metrics. average resulting system datasets yielded EERs 7.19, 16.85, 11.51, 11.90% HMM-DNN, DNN-HMM, DNN-GMM, GMM-DNN, respectively. Furthermore, found DNN-GMM least computational complexity compared to both talking Conversely, required greatest amount training time. Findings EER AUC values depended database when comparing performances.