Hierarchical Multi-Task Learning Based on Interactive Multi-Head Attention Feature Fusion for Speech Depression Recognition
IEEE Access · 2025 · 13 · 51208-51219
DOI: 10.1109/access.2025.3551549 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2025. Verfasst von Yujuan Xing und 3 weiteren Autor:innen. Die Arbeit wurde bisher 8-mal zitiert (Crossref) und nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Natural Science Foundation of China, Ph.D. Program of Yanyuan Science and Technology Innovation Fund.
Themen: Depression
Abstract (Originalsprache)
Depression, a severe psychiatric disorder with potential implications for suicide, has garnered increasing attention from researchers and the general public alike. Due to the accessibility of speech data, researchers have proposed many automatic depression recognition methods based on speech. Extracting effective features and designing an efficient recognition framework pose challenges in speech-based depression recognition. This study proposed a hierarchical multi-task learning framework based on interactive multi-head attention feature fusion (HMTL-IMHAFF), aiming to enhance the depression recognition accuracy by effectively leveraging the emotional and gender-related information embedded in speech. Firstly, we proposed a feature fusion approach based on interactive multi-head attention (IMHAFF) to comprehensively investigate the interplay among emotion valence, gender, and depression in order to obtain enhanced feature representation for depression recognition. Secondly, we developed a hierarchical multi-task learning framework with two levels. The first level consisted of sub-models related to emotion valence and gender to obtain the primary recognition results. The second level served as the final decision-making stage, which included a connection layer followed by a softmax layer that received input from the outputs of the first level. In addition, HMTL-IMHAFF used Long short-term memory (LSTM) model to capture the temporal relationship of speech feature sequences. Finally, the proposed method was comprehensively tested on the AVEC2014 dataset, and good results were achieved. Compared with the current related methods, the proposed method also stood out, obtaining an accuracy of 83.02%, which was 15% higher than the baseline method.
Förderung
- National Natural Science Foundation of China
- Ph.D. Program of Yanyuan Science and Technology Innovation Fund
Zitieren
APA
Xing, Y., He, R., Zhang, C., & Tan, P. (2025). Hierarchical Multi-Task Learning Based on Interactive Multi-Head Attention Feature Fusion for Speech Depression Recognition. IEEE Access, 13, 51208-51219. https://doi.org/10.1109/access.2025.3551549
Vancouver
Xing Y, He R, Zhang C, Tan P. Hierarchical Multi-Task Learning Based on Interactive Multi-Head Attention Feature Fusion for Speech Depression Recognition. IEEE Access. 2025;13:51208-51219. doi:10.1109/access.2025.3551549
BibTeX
@article{Xing2025,
author = {Xing, Yujuan and He, Ruifang and Zhang, Chengwen and Tan, Ping},
title = {Hierarchical Multi-Task Learning Based on Interactive Multi-Head Attention Feature Fusion for Speech Depression Recognition},
journal = {IEEE Access},
year = {2025},
volume = {13},
pages = {51208--51219},
doi = {10.1109/access.2025.3551549}
}
EUVIMED