Complex-Domain Decomposition with Bidirectional Convolutions Model for Speech Emotion Recognition

This study contributes to the domain of speech emotion recognition (SER) by proposing a novel framework based on a dual-path convolutional architecture for emotion classification. The framework integrates a Two-Stream Convolutional Block (TSB) and a Linear Recurrent Block (LRB) to jointly model localized spectral patterns and long-term temporal dependencies in speech segments. Evaluations across diverse datasets showcase the model's enhanced performance relative to current approaches. The TSB is designed to isolate discriminative emotion cues from time-frequency representations, while the LRB leverages sequential dynamics through stable recurrence. The proposed framework attains pioneering performance levels across critical evaluation dimensions, spanning both Unweighted Accuracy (UA) and Weighted Accuracy (WA). A notable advancement lies in the model's improved recognition of neutral emotions, addressing a persistent challenge in SER research. These outcomes underscore the efficacy of the introduced methodology in elevating SER performance.