Revamping EmoTech: Leveraging Dilated Convolution for Cutting-Edge Robust Speech Emotion Recognition

Emotion recognition plays a crucial role in enhancing human-computer interaction by enabling systems to respond naturally to users’ emotions. Traditional approaches to speech emotion recognition (SER) often relied on manual feature extraction and conventional machine learning models, which struggled to capture complex emotional patterns and long-range dependencies in speech. This paper presents a robust SER framework that addresses these limitations by leveraging advanced feature extraction techniques and innovative architectural elements. The proposed model integrates dilated convolutional layers with Bidirectional Long Short-Term Memory Networks (BiLSTMs) to effectively capture long-range dependencies and intricate emotional patterns in speech signals. Using Mel Frequency Cepstral Coefficients (MFCCs) as the primary features, the network processes data through two parallel pathways: one employing a BiLSTM network for sequence modeling and the other utilizing dilated convolutions to extract temporal spectral features. To address data imbalance and enhance model robustness, data augmentation techniques such as pitch shifting and time-stretching are applied. The fused features are then fed into fully connected layers to classify emotions into 10 classes for IEMOCAP and 6 classes for CREMA-D. Experimental results on the IEMOCAP and CREMA-D datasets demonstrate the model’s effectiveness, achieving test accuracies of 87% and 86%, respectively. Incorporating dilated convolutions and data augmentation significantly improves the recognition of subtle emotional cues, offering a reliable and scalable solution for speech emotion recognition tasks.