Mental Health Detection from Speech Signal: A Convolution Neural Networks Approach

Abstract

Mental health disorder is a global topic, the current situation is particularly serious in China. The objective and automated detecting of mental health using speech signal has become popular. In the absence of depressed speech corpus, the authors regard depression as a negative emotion, and build the model by Convolution Neural Networks (CNNs), a machine learning method for detecting mental health disorder interchanging with emotional speech. In this experiment, the segmented speech was represented as a spectrogram in the frequency-time domain via a Short-Time Fourier Transform (STFT), and these images were as input of the CNNs model. It highlights some advantages that CNNs can offer mental health detection. Results indicate that it is a good attempt and this method can be directly utilized by interchanging with emotional speech.