Hybrid GNN-LSTM: Enhancing Speech Emotion Recognition

Speech recognition systems can convert spoken language into text with remarkable accuracy. To recognize emotion in a speech means enabling further human-computer interaction. Traditional approaches for speech emotion detection use SVMs, CNNs, and HMMs. This study proposes a novel approach combining Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) architectures as the neural network aggregator to classify emotions from the SAVEE dataset. Using features extracted with OpenSmile. Without any advanced data augmentation, only padding and truncation applied, the proposed GNN-LSTM model achieved an accuracy of 81.7%, outperforming previous methods applied to the SAVEE dataset.