Biologically-Inspired Speech Emotion Recognition Using Rate Map Representations: An Application to the ShEMO Persian Speech Database

Abstract

This paper presents an innovative Speech Emotion Recognition (SER) model, inspired by the human auditory system, for analyzing and interpreting emotions in speech. Our proposed model utilizes a rate map representation to encode the spectro-temporal characteristics of auditory nerve activity, closely mimicking the intricate processes of human auditory perception. This model comprises several stages: pre-emphasis of the audio signal, cochlear filtering using a Gammatone Filter bank (GTF), neuromechanical transduction modeled by the Dau inner hair cell model, and the assembly of a rate map representation through integration of responses from each frequency channel. We apply this model to the ShEMO database, an extensive collection of Persian emotional speech, to detect and classify a spectrum of emotions. Our experimental results, obtained using deep learning architectures, demonstrate the effectiveness of the proposed model. We report the highest classification metrics with the Mobilenet architecture, achieving a performance of 71.57% and an F1 score of 51.52%. Overall, this study contributes to the field of speech emotion detection by offering a biologically-inspired model, validated with a substantial dataset, and yielding promising results in emotion classification using advanced machine learning techniques.