Application of cellular neural networks with the recurrent Perceptron learning algorithm for speech emotion recognition

Abstract

The paper proposes a novel approach to speech emotion recognition using Cellular Neural Networks (CeNNs) with the Recurrent Perceptron Learning Algorithm (RPLA). The model classifies emotions into two categories: positive and negative, based on audio signals. Experiments conducted on a combined dataset of four original databases (EmoDB, SAVEE, TESS, CREMA-D) with 10,257 samples show that a five-layer CeNNs model achieves an accuracy of 82% ± 0.02 (p = 0.0001 compared to Transformer), outperforming Gaussian Mixture Models (GMM, 68%), Support Vector Machines (SVM, 72%), Long Short-Term Memory networks (LSTM, 75%), and Transformers (80%). An average processing latency of 50 ms supports real-time applications. This research enhances human-machine interaction in virtual assistants, customer service, and mental health support.