Deep scattering network for speech emotion recognition

Abstract

This paper introduces scattering transform for speech emotion recognition (SER). Scattering transform generates feature representations which remain stable to deformations and shifting in time and frequency without much loss of information. In speech, the emotion cues are spread across time and localised in frequency. The time and frequency invariance characteristic of scattering coefficients provides a representation robust against emotion irrelevant variations e.g., different speakers, language, gender etc. while preserving the variations caused by emotion cues. Hence, such a representation captures the emotion information more efficiently from speech. We perform experiments to compare scattering coefficients with standard melfrequency cepstral coefficients (MFCCs) over different databases. It is observed that frequency scattering performs better than time-domain scattering and MFCCs. We also investigate layerwise scattering coefficients to analyse the importance of time shift and deformation stable scalogram and modulation spectrum coefficients for SER. We observe that layer-wise coefficients taken independently also perform better than MFCCs.

Multi-Task Conformer with Multi-Feature Combination for Speech Emotion Recognition

“…On the contrary, dimensional emotion can represent human emotion in a wider range than that of categorical emotion. In addition, distinguishing categorical emotion tends to cause confusion if arousal and valence levels are similar [6,7]. Therefore, in this study, we focus on the arousal and valence of dimensional emotion, and, in particular, on discrete arousal and valence tasks, since arousal and valence recognition can be designed as a regression task [8][9][10] or as a categorical task [11][12][13].…”