Are You Speaking with a Mask? An Investigation on Attention Based Deep Temporal Convolutional Neural Networks for Mask Detection Task

Abstract

When writing this article, COVID-19 as a global epidemic, has affected more than 200 countries and territories globally and lead to more than 694,000 deaths. Wearing a mask is one of most convenient, cheap, and efficient precautions. Moreover, guaranteeing a quality of the speech under the condition of wearing a mask is crucial in real-world telecommunication technologies. To this line, the goal of the ComParE 2020 Mask condition recognition of speakers subchallenge is to recognize the states of speakers with or without facial masks worn. In this work, we present three modeling methods under the deep neural network framework, namely Convolutional Recurrent Neural Network(CRNN), Convolutional Temporal Convolutional Network(CTCNs) and CTCNs combined with utterance level features, respectively. Furthermore, we use cycle mode to fill the samples to further enhance the system performance. In the CTCNs model, we tried different network depths. Finally, the experimental results demonstrate the effectiveness of the CTCNs network structure, which can reach an unweighted average recall (UAR) at 66.4% on the development set. This is higher than the result of baseline, which is 64.4% in S2SAE+SVM nerwork(a significance level at \\(p < 0.001\\) by one-tailed z-test). It demonstrates the good performance of our proposed network.