Voice-Derived Speech Features of Anhedonia in Adults With Depression: Exploratory Secondary Analysis of a Semistructured Clinical Interview

Abstract

Background: Anhedonia is a core, clinically consequential feature of depression, and its motivational and behavioral features may be difficult to observe through self-report alone. Speech is an effortful, dynamic, and socially contingent behavior that may index the motivational disruptions central to anhedonia, but few studies have examined anhedonia specifically using a broad, multimodal set of voice features. Objective: We conducted an exploratory, hypothesis-generating analysis of whether voice-derived speech features from a semistructured clinical interview are associated with concurrent anhedonia severity and with anhedonia 4 weeks later, adjusted for baseline severity. Methods: In a secondary analysis of a depression intervention trial, we analyzed recorded 17-item Hamilton Rating Scale for Depression interviews from adults with depression (anhedonia: n=44; depression severity: n=45) collected before participants initiated a behavioral intervention. Of these, 36 provided symptom data at week 4. Acoustic (pitch, spectral or mel-frequency cepstral coefficients, and jitter or shimmer), prosodic, linguistic (lexical richness and sentiment), topic-model, and automatic speech recognition quality features were extracted. After reducing the high-dimensional acoustic set to 9 theory-driven domain composites, 77 features were tested per outcome. Anhedonia (Mood and Anxiety Symptom Questionnaire, Anhedonic Depression Scale) and general depression severity (Quick Inventory of Depressive Symptomatology, Self-Report [QIDS-SR]) were examined with univariate Pearson correlations, with significance assessed by permutation tests and interpreted against a conservative P=.01 threshold, given the exploratory, multiple-testing context. For the week-4 analyses, baseline voice features were correlated, with week-4 symptom scores residualized for baseline symptoms and treatment condition. We report unadjusted P values and emphasize effect sizes, given the exploratory aims. Results: Higher baseline anhedonia was associated with briefer, less redundant speech: higher hapax ratio and lexical diversity (r=0.46; P=.003) alongside fewer content and total words (rs≈–0.43; P≤.01) and shorter total speaking duration (r=–0.41; P=.007), all exceeding the P=.01 threshold. Weaker, subthreshold trends included more negative sentiment (compound r=–0.34; P=.02), less work or school topic content (r=–0.31; P=.04), and flatter pitch variation (fundamental frequency SD r=–0.29; P=.06). General depression severity (QIDS-SR) showed no associations that exceeded the threshold. In baseline-adjusted week-4 analyses, no feature was associated with anhedonia at P=.01, whereas lower baseline voice perturbation (jitter composite; r=–0.46; P=.005) was associated with higher-than-predicted week-4 depression severity. Conclusions: Higher anhedonia aligned most robustly with a voice signature of reduced speech output and sparser but proportionally more diverse speech, with weaker, subthreshold trends toward flatter prosody and more negative sentiment, consistent with diminished behavioral engagement rather than uniform flattening of expressivity. Because the sample was small, predominantly female, and drawn from a depressed sample, and because many features were tested, findings are preliminary and require replication with larger samples and more naturalistic speech.