Emotionally engaged speech reveals acoustic markers of depression and suicide risk.

BACKGROUND: Major depressive disorder exists along a continuum, from health through remission to active depression. Differentiating these states remains challenging, and suicide risk may not be fully captured by self-report. Objective markers, such as speech, which integrates affective, cognitive, and motor processes, offer a promising avenue for fine-grained state differentiation. METHODS: Ninety-nine participants (healthy controls, depressive remission, and depressive episode) completed word-reading, question-answering, and story-reading tasks with positive, neutral, and negative emotional content. Twenty-three acoustic features were extracted and modeled using Long Short-Term Memory networks for three-class depression classification and binary suicide risk identification. Five-fold cross-validation and Shapley Additive Explanations were used for performance evaluation and model interpretation. RESULTS: Speech-based modeling reliably distinguished depressive states, with remission occupying an intermediate acoustic profile. Question-answering tasks achieved the highest accuracy for depression (75.89%) and suicide risk (78.03%). Emotional materials enhanced group separation, with positive and negative valence outperforming neutral speech. Feature analysis highlighted Energy, Teager Energy, and Spectral Flatness for healthy/low-risk profiles, and PitchTrack, Zero-Crossing Rate, and specific MFCCs for depressive/high-risk profiles. CONCLUSIONS: Emotionally engaging speech tasks enable interpretable, fine-grained differentiation of depressive states and support suicide risk identification. Depressive remission presents as an acoustically distinct intermediate state rather than a full return to normative patterns, emphasizing the value of affective speech paradigms for scalable mental health assessment.