Scalable, non-invasive depression monitoring with smartphone speech: a multimodal benchmark and topic analysis

Abstract Semantic patterns in weekly voice diaries provide a scalable window onto depression. We collected 3,151 weekly free-speech diaries from 284 German-speaking adults (156 controls, 128 with major depressive disorder) using the ReMAP smart-phone app. Automatic speech recognition produced transcripts; lexical, prosodic and openSMILE acoustics, plus sentence embeddings from the 8-billion-parameter Qwen3-8B model, were extracted. Support-vector regression with nested crossvalidation predicted Beck Depression Inventory (BDI) scores. Qwen3-8B embeddings substantially outperformed all classical features, achieving a mean absolute error of 4.56 points and R 2 = 0.34. The best-performing model combined embeddings with lexical and prosodic metrics (MAE = 4.44, R 2 = 0.35), representing only a modest improvement over embeddings alone. BERTopic applied to the embeddings revealed ten themes; depressive severity varied strongly between themes (Kruskal–Wallis H = 663.9, p < 0.001) with “Persistent Low Mood” and “Pain Distress” most associated with high BDI scores (Cliff’s δ ≤ 0.99). These results show that modern language model embeddings capture the primary signal of depression severity in everyday speech, supporting privacy-preserving, non-invasive mental-health monitoring at population scale.