Speaker Identification for Disguised Voices Based on Modified SVM Classifier
Noor Ahmad Al Hindawi, Ismail Mohd Adnan Shahin, Ali Bou Nassif · 2021
Since voice disguise forms a significant threat in the plethora of illegal applications, it is essential to be able to identify the unknown speaker. This work focuses on scheming a modified Support Vector Machine (SVM) as a classifier to enhance the degraded speaker identification performance for disguised voices under an extreme high-pitched condition in a neutral talking environment. This research utilizes three different speech datasets: Arabic Emirati-accented database, “Speech Under Simulated and Actual Stress” (SUSAS) English database, and “Ryerson Audio-Visual Database of Emotional Speech and Song” (RAVDESS) English database. Our results show that modified SVM reports an average speaker identification performance for disguised voices equal to 93.95%. Our work demonstrates that modified SVM is superior to other classical classifiers such as: K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP), Radial Basis Function (RBF), Naïve Bayes (NB), and the conventional SVM.