Identification of depressive symptoms through acoustic analysis of the voice using convolutional neural networks

According to the WHO, depression is a common mental disorder characterized by persistent sadness and loss of interest in activities that were previously enjoyed. In 2020, the WHO estimated that more than 350 million people in the world presented it. In extreme cases, this disorder has been linked to suicide. Recent studies have shown that the use of the voice can serve as a promising biomarker for the detection of depression. This work addresses the effectiveness of using a machine learning model, particularly convolutional neural networks, to be able to detect symptoms of minor and major depression through voice. This model is trained with log scale spectrograms based on the audio recordings of 52 participants. The recordings were classified into two groups (low and high depressive symptoms, respectively), where each group contained 26 participants. Each voice recording was obtained through a determined reading of a questionnaire through a web platform. The classification of these groups was carried out according to an average between the BDI, CESD, CRSD and PHQ-9 measurement scales, provided by the same platform. Preprocessing, random polling, and sliding windows have allowed each audio to be significantly normalized. The results of the training of this model are presented in this work.