Anti-Aliased Convolutional with Data Augmentation for Speech Emotion Recognition

In this work, we propose a comprehensive framework for speech emotion recognition that combines advanced data augmentation techniques and an innovative deep learning architecture to enhance model performance. Our data augmentation strategies, including noise addition, time stretching, pitch shifting, and time shifting, significantly diversify the training dataset, leading to improved generalization and robustness against real-world variations in speech. We introduce a novel deep learning model architecture based on Anti-Aliased Convolutional (AA-Conv1D) layers, which preserves spatial hierarchies and enhances the model's ability to capture fine-grained emotional nuances in speech signals. Experimental results demonstrate that our model outperforms state-of-the-art methods, achieving a weighted accuracy of 79.38%, which surpasses the performance of existing models, SepTr + LeRaC and EMix-NS.