情绪识别
语音识别
计算机科学
心理学
适应(眼睛)
人工神经网络
人工智能
特征(语言学)
情绪分类
认知心理学
语音处理
自然语言处理
说话人识别
情感计算
深层神经网络
作者
Yan Zhao,Jincen Wang,Cheng Lu,Sunan LI,Fan Liu,Björn Wolfgang Schuller,Wenming Zheng,Yuan Zong
标识
DOI:10.1109/taffc.2026.3746232
摘要
Speech emotion recognition (SER) remains challenging when the training (source) and testing (target) data are drawn from different speech corpora, a setting commonly referred to as cross-corpus SER. Although domain adaptation (DA) methods show promise for mitigating this domain shift, many conventional unsupervised DA methods require access to labeled source speech data, limiting their applicability in privacy-sensitive scenarios, e.g., patient recordings from different hospitals. To address this limitation, we investigate a stricter setting of cross-corpus SER, i.e., source-free cross-corpus SER, aligned with the source-free DA (SFDA) paradigm. In this setting, no source speech data are accessible and only a pre-trained source SER model is available. Building on our prior ECAN (emotion-aware contrastive adaptation network), we propose E$^{2}$CAN, an enhanced emotion-aware contrastive adaptation network tailored to source-free cross-corpus SER. To evaluate E$^{2}$CAN, we establish an evaluation protocol on five widely-used speech corpora: CASIA, EmoDB, EMOVO, eNTERFACE, and IEMOCAP, and compare against seven recent state-of-the-art SFDA methods. Experiments demonstrate the effectiveness and superiority of E$^{2}$CAN in addressing source-free cross-corpus SER.
科研通智能强力驱动
Strongly Powered by AbleSci AI