A Comparative Performance of Optimizers and Tuning of Neural Networks for Spoof Detection Framework

Abstract

The breakthroughs in securing speaker verification systems have been challenging and yet are explored by many researchers over the past five years. The compromise in security of these systems is due to naturally sounding synthetic speech and handiness of the recording devices. For developing a spoof detection system, the back-end classifier plays an integral role in differentiating spoofed speech from genuine speech. This work conducts the experimental analysis and comparison of up-to-date optimization techniques for a modified form of Convolutional Neural Network (CNN) architecture which is Light CNN (LCNN). The network is standardized by exploring various optimizers such as Adaptive moment estimation, and other adaptive algorithms, Root Mean Square propagation and Stochastic Gradient Descent (SGD) algorithms for spoof detection task. Furthermore, the activation functions and learning rates are also tested to investigate the hyperparameter configuration for faster convergence and improving the training accuracy. The counter measure systems are trained and validated on ASV spoof 2019 dataset with Logical (LA) and Physical Access (PA) attack data. The experimental results show optimizers perform better for LA attack in contrast to PA attack. Additionally, the lowest Equal Error Rate (EER) of 9.07 is obtained for softmax activation with SGD with momentum wrt LA attack and 9.951 for SGD with nestrov wrt PA attack.