Development and Analysis of Convolutional Neural Network based Accurate Speech Emotion Recognition Models

Automatic speech recognition is a major topic in artificial intelligence and machine learning, with the intent of developing machines that can communicate with humans through speech. Recently, with the emergence of the deep-learning paradigm, end to-end models that extract features and train directly from the raw speech signal have been developed. With the goal of more precisely classifying emotions from speech, this paper presents a comparative analysis on two deep-learning architectures that improves on the models available in the literature in terms of accuracy. Using a combined CNN-LSTM architecture and a CNN-Transformer encoder architecture, this work analyses the complete deep learning strategy for extracting distinct spatial and temporal features and classifying the emotions from speech. Experiments are carried out on the RAVDESS dataset. The CNN-Transformer encoder network achieves high accuracy 82% in these networks, while the CNN-LSTM network achieves 74%.