Low-burden speech-based screening of depressive symptoms using ultra-short speech samples

Speech holds promise for depressive symptom screening, but practical applications require clearer guidance on effective segmentation, optimal speech tasks, minimum duration, and feature-model matching.

To enable low-burden screening for depressive symptoms, this study aims to evaluate the effectiveness of short speech samples and systematically investigate how to identify the optimal practical combination of different speech tasks, time-window lengths, and the matching between features and models.

Using the HMU-DDAC-24 dataset of 143 participants aged 18-22 years, we evaluated 10 segmentation lengths (0.1-15 s), 5 models (Random Forest, Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Transformer, Extended Long Short-Term Memory (xLSTM)), and 3 feature types (frame-level, statistical, fused). We also assessed 19 window lengths (1-35 s) across four speech tasks and their combinations. In addition, supplementary cross-dataset validation analysis was performed using the external Emotional Audio-Textual Depression (EATD) dataset.

Frame-level features performed best with sequence models and short windows (0.1-2 s; Transformer accuracy: 0.731 at 2 s). Statistical features paired best with tree models and short windows (Random Forest accuracy: 0.698 at 1 s). Fused features improved tree models (Random Forest accuracy: 0.703 at 4 s) but introduced noise into sequence models. Task 4 (positive recall) achieved the best overall performance (Area Under the Receiver Operating Characteristic Curve (AUC) = 0.860), outperforming all single and combined tasks, showing no multi-task synergy. A 1-s segment performed comparably to full-length speech (3.14% relative performance loss). Kullback-Leibler divergence analysis showed short segments closely approximated the full recording's distribution (Task 4 KL = 9.56 at 1 s).

When combined with appropriate speech tasks and feature-model matching, short speech samples can facilitate low-burden screening for depressive symptoms. Supplementary analysis on the EATD dataset supported these trends, though absolute performance depended on feature representation. We recommend positive recall with 2-s truncation, frame-level features, and Transformer for accuracy-oriented settings; and 1-s truncation, statistical features, and Random Forest for resource-limited settings. These findings may inform rapid, remote speech-based screening for depressive symptoms, especially in resource-limited primary-care settings.