文献类型：期刊文献

英文题名：TFCTL: Time-Frequency Calibrated Transfer Learning for cross domain depression detection

作者：Li, Dongdong[1];Ding, Li[1];Yang, Zuo[1];Wang, Zhe[1];Zhao, Ke[2]

机构：[1]East China Univ Sci & Technol, 130 Meilong Rd, Shanghai 200237, Peoples R China;[2]Wenzhou Med Univ, Wenzhou 325000, Peoples R China

年份：2025

卷号：162

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：;EI(收录号：20253919234356);WOS:【SCI-EXPANDED(收录号:WOS:001584926300001)】；

基金：Acknowledgments This work was supported by the National Natural Science Foundation of China under Grant No. 62276098 and the Wenzhou Science and Technology Bureau under Grant No. ZS2024001.

语种：英文

外文关键词：Frequency-Delay Neural Network; Time-frequency features; Transfer learning; Cross-domain adaptation; Depression detection

摘要：As the application of speech-based depression detection systems expands, differences in cross-domain data distribution pose significant challenges. This paper proposes the framework of Time-Frequency Calibrated Transfer Learning (TFCTL). This framework first introduces Frequency-Delay Neural Network (FDNN), inspired by Time-Delay Neural Network (TDNN), extending the concept of temporal feature extraction using sliding windows and weight sharing from the time domain to the frequency domain. A multi-level information aggregation module then integrates features of varying abstraction levels from both time-delay and frequency-delay neural networks, balancing global and local speech information. Finally, TFCTL uses transfer learning to calibrate the distribution of the aggregated time-frequency embedding vectors, uncovering commonalities of depression features across different domains. Cross-speaker and cross-corpus experiments were conducted using the Chinese Multimodal Depression Corpus (CMDC) and the Distress Analysis Interview Corpus Wizard-of-Oz (DAIC-WOZ). In cross-speaker scenarios, TFCTL achieved F1 scores of 0.7324 on DAIC and 0.9660 on CMDC, outperforming other methods. In cross-corpus scenarios, TFCTL achieved F1 scores of 0.6743 on DAIC and 0.6879 on CMDC, demonstrating its robustness in addressing domain mismatch issues. The source code used in the paper is available at https://anonymous.4open.science/r/TFCTL-A545/.