Аңдатпа
The central difficulty of cross-corpus speech emotion recognition lies not only in acoustic mismatch between datasets, but also in the fact that different corpora encode emotion through different recording conditions, speaking styles, and class-conditional structures. This paper presents a controlled empirical study of directional asymmetry in domain-adversarial cross-corpus speech emotion recognition using three public English emotional speech corpora, namely IEMOCAP, MSP-Podcast, and CREMA-D. Under a unified four-class setting, we compare mean pooling, attention pooling, and chunk-based aggregation, and then examine two progressively stronger adaptation regimes built on the chunk model: utterance-level adversarial alignment and utterance+chunk-level alignment. The results show that transfer behavior is clearly direction-dependent. In the acted-to-natural setting, utterance-level alignment yields a small improvement, while bootstrap analysis shows that the UAR gain is modest and should be interpreted cautiously. Additional no-excited sensitivity experiments confirm that this trend is not merely an artifact of merging IEMOCAP excited into happy. In the opposite natural-to-acted direction, the plain chunk baseline remains stronger than the adaptation variants, confirming that alignment is not uniformly beneficial. We further compare the adversarial variants with MMD and CORAL alignment baselines; MMD is competitive with utterance-level adversarial alignment in PODCAST-target settings, whereas CORAL provides no meaningful gain. Additional evidence from class-wise recall redistribution, confusion analysis, and representation-space visualization suggests that stronger domain invariance does not necessarily preserve emotion-discriminative structure. Within the examined corpus combinations and frozen wav2vec 2.0 Base setting, these findings support direction-aware evaluation and selective adaptation rather than the assumption that stronger alignment is uniformly beneficial.