A feature selection based parallelized CNN-BiGRU network for speech emotion recognition in Odia language

Emotion recognition from speech is an integral part of human interaction. This paper represents work that includes the creation and evaluation of speech emotion recognition on the Odia database. Previous research in this field implemented several benchmark datasets. In this work, we use two benchmark datasets for cross-validation with our own created Odia dataset name as SITB-OSED. Initially, the spectral, prosodic, and voice quality features are extracted from a raw audio file; secondly, a Gradient Boost Decision Tree (GBDT) feature selection method is used to remove all the redundant features and select the potential features. Here, two distinct series of experiments are performed. Firstly, the baseline model, which takes all the combined selected features, is chosen as input (spectral, prosodic, and voice quality features). Secondly, the proposed model processes all the selected features through two separate channels, the Convolutional neural network (CNN) and Bi-directional gated recurrent units (Bi-GRU). Specifically, the proposed method achieved 6.67%, 6.03%, and 5.55% higher accuracy than the baseline model on the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) Interactive Emotional Dyadic Motion Capture (IEMOCAP) and SITB-OSED datasets. We also report that the proposed parallelized CNN-BiGRU model outperforms the recent state-of-the-art methods with an accuracy of 82.29% and 78.54% on the RAVDESS and IEMOCAP datasets, respectively. For our SITB-OSED dataset, the overall recognition accuracy of 84.02% is achieved.