Enhanced Speech Based Human Emotion Identification by Hybrid Deep Learning and Optimal Features with Meta-Heuristic Algorithm

Speech Emotion Identification (SEI) is a difficult job in emotional computing that faces challenges with conventional acoustic methods. To address this, a Hybrid Deep Learning (HDL) model is proposed in this research. SEI datasets are collected and pre-examined by artifact removal and filtering techniques. The extracted features include, Tonal Power(TP), spectral flux(SF), Mel Scale Spectrogram(MSS) and Mel Frequency Cepstral Coefficient(MFCC). Optimal feature selection is performed using the Adaptive Search based Deer Hunting (AS-DH) algorithm to improve learning performance. The HDL model combines Recurrent Neural Network (RNN) and Deep Neural Networks (DNN) enhanced by AS-DH, achieving high accuracy in classifying affections like happy, sad, fear, calm and anger. The proposed AS-DH-HDL method outperforms various existing HDL methods with improvement in accuracy at a learning rate of 85. The performance evaluation demonstrates the superior performance of the developed model across various metrics.