Beyond Short-Frame Acoustic Features: Capturing Long-Term Speech Patterns for Depression Detection

Speech-based depression detection is promising for objective mental health assessment. However, conventional methods relying on short-frame acoustic features often fail to capture long-term temporal and behavioral characteristics of speech essential for modeling depression-specific speaking patterns. Herein, four novel acoustic feature sets extracted from long-term speech are proposed: utterance interval feature set (UIFS), pause interval feature set (PIFS), response interval feature set (RIFS), and speech density (SD). These features explicitly characterize temporal structures and session-level speech behaviors beyond short-frame analysis. These features are combined with conventional acoustic features, including standard features extracted using openSMILE and voice level features, and evaluated using support vector machines under subject-independent conditions for the binary classification of depressed and nondepressed speakers. Incorporating the proposed features improves classification performance compared with baseline features (accuracy: 0.54 for openSMILE and 0.52 for openSMILE + voice level features). The configuration integrating all four proposed feature sets achieves an accuracy of 0.58, a precision of 0.56, a recall of 0.58, and a specificity of 0.58, indicating consistent performance gains under subject-independent and strictly controlled evaluation conditions. Thus, depression-related speech patterns can be captured by explicitly modeling temporal and behavioral speech characteristics across entire dialog sessions. This study contributes to advancing acoustic feature design for speech-based depression detection and developing clinically supportive screening and monitoring technologies.