Enhancing Speech Emotion Recognition through a Cross-Dataset Analysis: Exploring Improved Models

In recent years, there has been a growing interest in Speech Emotion Recognition (SER) due to its wide-ranging applications. To enhance the accuracy of emotion recognition systems, this paper employs a combination of convolutional neural networks (CNN) and long short-term memory (LSTM), both of which have been previously proposed and extensively studied in various research works. The ConvLSTM model adeptly captures important audio features effectively, leading to improved performance. The study includes a comprehensive analysis across multiple datasets, including RAVDESS, SAVEE, EMO-DB, and TESS, to evaluate the performance of the ConvLSTM model. Data augmentation (DA) techniques are employed to increase the training data, and feature selection using the Analysis of Variance (ANOVA) technique is utilized to identify the most informative features. The proposed approach demonstrates its effectiveness in identifying optimal configurations, which leads to high accuracy rates. Specifically, the approach achieves an accuracy of 97.16% for the the RAVDESS dataset, 97.45% for SAVEE, 97.98% for EMODB, and a perfect accuracy of 100% for TESS. Furthermore, when combining the RAVDESS, SAVEE, and EMODB datasets, the approach achieves an accuracy of 97.37% for the R+T+E dataset, and an even higher accuracy of 99.27% when combining the RA VDESS, SA VEE, and TESS datasets (R+S+T). These results underscore the effectiveness of the proposed approach in accurately classifying speech emotions across diverse datasets, highlighting its potential for practical applications in emotion recognition tasks.