Attention-Based Neural Networks for Speech Emotion Recognition: A Comparative Study of FCN, BiLSTM, and LSTM Models

Speech Emotion Recognition is considered a pivotal technology in human-computer interaction, facilitating the systems to detect and react to emotional states. Despite advancements in SER methods, the challenge of accurately labeling emotions persists due to the intricate nature of emotional features. In this paper, we aim to enhance accuracy, to achieve this, a model combining Mel-Frequency Cepstral Coefficients (MFCC) with three advanced deep learning architectures: Fully Convolution Networks (FCN) and Recurrent Neural Networks (RNNs), which include Long Short-Term Memory (LSTM) And Bidirectional Long Short-Term Memory (BiLSTM) networks is proposed to recognize four emotions, happy, angry, sad and neutral. The model is trained, validated, and tested using an extensive dataset, created by merging RAVDEES and TESS datasets, leading to significant improvement in performance, achieving an accuracy of 95% for the attention-based FCN, 92% for attention-based RNN-LSTM, and 93% for attention based RNN-BiLSTM. This research advances the field of SER and emphasizes the potential of attention mechanisms in solving intricate classification of tasks in speech analysis.