MBDA: A Multi-scale Bidirectional Perception Approach for Cross-Corpus Speech Emotion Recognition

Effectively combining context to extract emotion-related features poses a significant challenge in the task of speech emotion recognition (SER). To address this challenge, this paper proposes the Bidirectional Temporal Multi-Scale Attention Network (BTMA). Our research combines varying receptive field with self-attention, automatically focusing on key information across different scales. Simultaneously, the bidirectional time series ensures the comprehensive extraction of emotional features in speech. In order to achieve cross-corpus training and testing, BTMA is integrated with an adversarial discriminant domain adaptation (ADDA) method, and further forms the Multi-scale Bidirectional Domain Adaptive Network (MBDA). This network effectively overcomes the differences among different corpora, providing more comprehensive domain-invariant features. Comparative experiments and ablation studies on 5 classic datasets demonstrate the superiority of our approach, showing substantial improvement in SER tasks.