Stress detection of children through speech signals in multi-speaker environment using deep learning

. Stress is a psychological problem that can affect anyone, including children. Detecting stress in children is a complex problem because they are generally unaware of the psychological problems they are experiencing and have verbal limitations that affect their communication skills with their parents. One of the biomarkers that can be used to detect stress is the voice (speech signal). The use of speech in stress detection has advantages in terms of convenience for the subject and ease of acquisition. This study proposes a stress detection model through speech signals in a multi-speaker environment. This model accepts audio input from the classroom environment, where there is noise and many speakers' voices overlap. The audio acquired is then separated using a speech separation algorithm based on an RNN architecture, producing output as segregated speech. The speech is then extracted for features and fed to the stress detection model based on CNN architecture, which predicts the speaker's stress status. The experimental results show that the proposed model is capable of speech separation with up to five speakers and predicts the stress status of the subject with an average accuracy of 95.6%.