Screening Trauma Through CNN-Based Voice Emotion Classification

Recently, modern people experience trauma symptom for various reasons. Trauma causes emotional control problems and anxiety. Although a psychiatric diagnosis is essential, people are reluctant to visit hospitals. In this paper, we propose a method for screening trauma based on voice audio data using convolutional neural networks. Among the six basic emotions, four emotions were used for screening trauma: fear, sad, happy, and neutral. The first pre-processing of adjusting the length of the audio data in units of 2 s and augmenting the number of data, and the second pre-processing is performed in order to convert voice temporal signal into a spectrogram image by short-time Fourier transform. The spectrogram images are trained through the four convolution neural networks. As a result, VGG-13 model showed the highest performance (98.96%) for screening trauma among others. A decision-level fusion strategy as a post-processing is adopted to determine the final traumatic state by confirming the maintenance of the same continuous state for the traumatic state estimated by the trained VGG-13 model. As a result, it was confirmed that high-accuracy voice-based trauma diagnosis is possible according to the setting value for continuous state observation.