Hybrid Architecture and pre-processing for Speech Emotion Recognition

Abstract

Accurate emotion recognition from speech is vital for improving human-computer interaction, particularly in advanced Mechatronic systems where responsive and adaptive behaviour is crucial. Traditional deep learning models like CNNs and LSTMs often need help with transient emotional nuances, and the presence of temporal relationships in audio signals limits their effectiveness in practical scenarios. This study introduces the Dendritic Convolutional LSTM (DCLSTM) architecture, which integrates dendritic computational principles to enhance the learning of complex spatial and temporal features in emotional speech. An advanced audio preprocessing pipeline is also implemented, systematically refining speech signals through noise reduction, spectral processing, and filtering to optimise model performance. By enabling more accurate and nuanced emotion recognition, this research represents a significant advancement in integrating human-like emotional understanding into Mechatronic systems, leading to greater effortless and flexible machine interactions.