Integrating LSTM and Transformer for Robust Voice Emotion Recognition

Accurate voice emotion recognition (VER) is crucial across various fields, including healthcare, customer service, and personalized dialogue systems. However, a significant challenge is the limited availability of datasets that encompass a wide range of emotional states across different language families. This paper presents a new model called the Hybrid-Module-Transformer, applied to the nEMO dataset, which is a valuable collection of emotional speech samples. The model integrates the strengths of both Transformer and LSTM architectures, resulting in exceptional performance in VER tasks. The nEMO dataset comprises over 3 hours of audio samples representing six distinct emotional states, providing a robust foundation for training and testing the model. Initial experiments show promising results, indicating the model's ability to accurately classify emotions.