Improved Cross-Corpus Speech Emotion Recognition Using Deep Local Domain Adaptation

Due to the scarcity of high-quality labeled speech emotion data, it is natural apply transfer learning recognition. However, learning-based recognition becomes more challenging because complexity and ambiguity emotion. Domain adaptation based on maximum mean discrepancy considers marginal alignment source domain target domain, but not pay regard class prior distribution in both domains, which results reduction efficiency. In order address problem, this study proposes a novel cross-corpus framework local adaption. A category-grained used evaluate distance between two relevant domains. According research findings, generalization ability model enhanced by using adaptive method. Compared with global non-adaptive methods, effectiveness significantly improved.