On the use of Deep Learning and Scattering Transform for Pathological voices recognition

Abstract

In the last few decades, Deep Neural Networks (DNNs) has shown outstanding performance in speech recognition applications. We demonstrate that the improved accuracy obtained by Deep Convolutional Neural Network (DCNN) arose from their capacity to extract discriminative representations which are robust to various sources of variability in speech signals. By this study, we propose a new algorithm, named Scattering Transform-Deep Convolutional Neural Network CNN: ST-DCNN to identify normal and pathological voices. The effectiveness of advances in speech features have been proven to be the root for an efficient pathological voices classification. The proposed algorithm involved two stages: First, scatter wavelet features are extracted. Then, DCNN is used to classify the voices samples. We evaluated the robustness of the proposed system in silent environments. The experimental results indicates that it achieves better performance with scattering wavelet and DCNN with the clean data within 99.62 % of recognition rate.