Adaptive Knowledge Fusion Model for Depression Recognition

In recent years, research on speech-based depression recognition has been growing rapidly. However, most existing methods adopt a single-perspective modeling approach, lacking an in-depth exploration of the multidimensional features of speech signals. As a result, they struggle to fully capture the subtle variations and key patterns in the speech of individuals with depression. To address this issue, this article proposes an adaptive knowledge fusion (AKF) model that effectively extracts time and channel domain features from speech signals, thereby enhancing the accuracy and robustness of depression recognition. Specifically, we propose a novel architecture featuring two specialized modules. The first, a time domain modeling module, holistically captures speech dynamics by hierarchically integrating features from local to global time scales. The second, a channel domain modeling module, enhances feature distinctiveness by promoting efficient cross-channel communication while simultaneously reducing redundancy within channel representations. Building upon these modules, we propose the AKF model, which employs a multipath modeling strategy to construct a unified framework that effectively integrates time and channel domain features. Experimental results on the androids corpus and clinical dataset demonstrate that the proposed model significantly outperforms all baseline models in depression recognition. Furthermore, through extensive additional experiments, we analyze the contribution of each submodule, further validating its effectiveness.