Speech Emotion Recognition using Time-frequency Random Circular Shift and Deep Neural Networks

Sylvain Xia (1) , Dominique Fourer (1) , Liliana Audin-Garcia (2) , Jean-Luc Rouas (2) , Takaaki Shochi (2)

1 IBISC - Informatique, BioInformatique, Systèmes Complexes
2 IMS - Laboratoire de l'intégration, du matériau au système

This paper addresses the problem of emotion recognition from a speech signal. Thus, we investigate a data augmentation technique based on circular shift of the input time-frequency representation which significantly enhances the emotion prediction results using a deep convolutional neural network method. After an investigation of the best combination of the method parameters, we comparatively assess several neural network architectures (Alexnet, Resnet and Inception) using our approach applied on two publicly available datasets: eNTERFACE05 and EMO-DB. Our results reveal an improvement of the prediction accuracy in comparison to a more complicated technique of the state of the art based on Discriminant Temporal Pyramid Matching (DCNN-DTPM).

Keywords: Speech Emotion Recognition (SER); Deep Convolutional Neural Networks; Time-frequency; Random Circular Shift (RCS)

Domains: Sound [cs.SD]; Artificial Intelligence [cs.AI]; Signal and Image Processing

Sylvain Xia, Dominique Fourer, Liliana Audin-Garcia, Jean-Luc Rouas, Takaaki Shochi. Speech Emotion Recognition using Time-frequency Random Circular Shift and Deep Neural Networks. Speech Prosody 2022, May 2022, Lisbon (Portugal), Portugal. pp.585--589, ⟨10.21437/SpeechProsody.2022-119⟩. ⟨hal-03583535⟩