Integrating Hybrid Deep Learning and Meta Heuristic Optimization for Enhanced Speech Emotion Recognition

Abstract

Emotion recognition has gained significant attention in recent years due to its importance in several fields like human-computer interaction, affective computing, and healthcare. Feature selection plays a vital role in improving the performance of emotion recognition from speech by identifying the most informative and relevant features. Extracting relevant features from speech signals plays a crucial role in accurately recognizing and classifying emotions. However, the high dimensionality and complexity of speech data pose challenges in selecting the most discriminative features. This thesis explores the development of an advanced Speech Emotion Recognition (SER) system utilizing a hybrid deep learning method enhanced by optimized heuristic algorithms to overcome limitations in conventional emotion recognition techniques. Emotion recognition in human-computer interaction is complex due to the intricate nature of speech features required for accurate classification. The proposed research addresses these challenges by combining acoustic and visual features, leveraging deep learning and optimized feature selection methods. Initially, SER datasets are collected from public sources and pre-processed to remove artifacts, ensuring signal clarity through noise filtering techniques. Feature extraction is then performed using Mel-Frequency Cepstral Coefficients, Mel scale spectrograms, tonal power, and spectral flux to capture nuanced aspects of speech signals. newlineTo further refine the model, optimal features are selected using the Deer Hunting with Adaptive Search (DH-AS) algorithm, enhancing the performance of hybrid deep learning model. The model comprises a Hybrid Deep Learning (HDL) approach, integrating Deep Neural Networks and newlineRecurrent Neural Networks, which is further tuned with DH-AS for improved emotion classification accuracy.