Classification Inconsistency Alignment Network for Cross-corpus Speech Emotion Recognition

Cross-corpus speech emotion recognition (SER) aims to transfer emotional information from a labeled source corpus to an unlabeled target corpus. Due to the characteristics of each corpus, models trained on source domain may classify target samples into incorrect categories, resulting in classification inconsistency problem. Therefore, this paper proposes classification inconsistency alignment network (CIAN). The core idea of CIAN is to construct two complementary and adversarial classifiers to optimize the decision boundary. Herein, primary classifier extracts discriminative emotional features by bridging source samples and corresponding labels. Then, the auxiliary classifier focuses on misclassified samples arising from domain discrepancy. Through the interaction of two classifiers, CIAN can identify all samples prone to inconsistent classification, thereby enhancing classification consistency while domain adaptation. Extensive experiments based on three public emotion corpora demonstrate that CIAN outperforms existing advanced methods in cross-corpus SER tasks.