This was an exploratory secondary analysis of speech recordings from a completed randomized clinical trial, comprising a cross-sectional component (baseline voice features and concurrent symptoms) and a longitudinal component (baseline voice features and week-4 symptoms adjusted for baseline). We report the study in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guideline for observational research [13]. A completed checklist is provided as Multimedia Appendix 1.

This report is a secondary analysis of speech recordings collected in a single-site randomized clinical trial that evaluated attention bias modification (ABM), a computerized cognitive training procedure, as an intervention for depression [14]. The parent trial was conducted at a large public university in the southern United States. Participants were recruited from the surrounding community and university through advertisements. Participants were eligible if they were 18 to 70 years of age, fluent in English, able to provide informed consent, and scored 13 or higher on the Quick Inventory of Depressive Symptomatology, Self-Report (QIDS-SR) at baseline. Those taking antidepressant medication were required to have had no change in medication or dose during the prior 12 weeks. Key exclusion criteria were active suicidal behavior or significant suicidal ideation within the prior 3 months (assessed with the Columbia-Suicide Severity Rating Scale); current or past bipolar or psychotic disorders and a current (past-12-month) alcohol or substance use disorder of at least moderate severity, both assessed with the Mini International Neuropsychiatric Interview; current opioid analgesic or systemic corticosteroid use; and ongoing psychotherapy [14]. The parent trial enrolled 145 adults. The present study and its voice-feature analyses were not part of the original trial design. Recordings of the clinician-administered symptom interview had been retained for interview-fidelity monitoring and were analyzed retrospectively for this report.

Voice features were extracted from the baseline (week 0) 17-item Hamilton Rating Scale for Depression (HRSD-17) interviews, which had been recorded for interview-fidelity monitoring. Eligibility for the present analyses required a baseline recording together with at least 1 usable baseline self-report outcome (Mood and Anxiety Symptom Questionnaire, Anhedonic Depression Scale [MASQ-AD] or QIDS-SR). We excluded 11 participants who had a recording but neither outcome, as they could not contribute to any association analysis (these 11 were also excluded from the parent-trial outcome analyses, with treatment assignment recorded as missing and no week-4 data). The analytic sample therefore comprised 45 participants. Within this sample, cross-sectional correlations used the participants with a valid baseline score on the relevant outcome (anhedonia [MASQ-AD]: n=44; depression severity [QIDS-SR]: n=45), and 36 also provided self-reported symptom data at week 4 and contributed to the baseline-adjusted longitudinal analyses. Because this subsample was small, the longitudinal analyses should be considered very exploratory, with statistical power to detect only large effects. The baseline interviews analyzed here were recorded between October 9, 2017, and April 10, 2019.

Anhedonia was assessed using the 10-item anhedonic depression scale from an adapted 30-item version of the Mood and Anxiety Symptom Questionnaire (MASQ-AD) [15]. The MASQ-AD emphasizes low positive affect and loss of interest or pleasure. Items are rated on a 1-5 Likert scale referencing the past week, and the 10 items are summed to yield a total score ranging from 10 to 50, with higher scores reflecting greater anhedonia. The MASQ-AD was the primary outcome for the present analyses.

Depression severity was assessed using the 16-item QIDS-SR [16]. The QIDS-SR is a brief self-report measure covering the 9 symptom criterion domains of the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) over the past 7 days. Items are scored on 0-3 anchors and summed to yield a total score from 0 to 27, with higher scores indicating greater depression severity. The QIDS-SR was included as a secondary, comparison outcome representing global depression severity.

Clinician-rated depression severity was assessed with the HRSD-17, a widely used semistructured interview that served as the source recording for voice-feature extraction. The HRSD-17 comprises 17 items assessing symptoms such as depressed mood, guilt, suicidality, insomnia, work and activities, psychomotor change, anxiety, and somatic symptoms. Items are scored on either 5-point (0-4) or 3-point (0-2) anchors depending on the symptom and summed to a total score ranging from 0 to 52, with higher scores indicating greater severity. The MASQ-AD, QIDS-SR, and HRSD-17 were each administered at baseline, and the MASQ-AD and QIDS-SR were readministered at week 4.

Voice features were extracted from the audio of the baseline HRSD-17 interview described earlier. The HRSD-17 is a brief semistructured interview that typically takes about 15-20 minutes to administer. Trained interviewers administered the HRSD-17, with weekly supervision by a licensed psychologist. Interrater reliability was high (intraclass correlation coefficient=0.98), and interviewers were blinded to participants’ intervention condition. Only participant speech (not interviewer speech) was used for participant-level feature summaries.

The parent trial was approved by the institutional review board of The University of Texas at Austin (protocol: 2016-06-0034), and all participants provided written informed consent for trial participation, including audio recording of the clinical interviews. The approved protocol and the informed consent obtained from participants explicitly provided for secondary analyses of these data. The present secondary analysis therefore fell within the scope of the original ethics approval and consent and did not require additional institutional review board review or additional participant consent. The present secondary analysis used these previously collected, approved recordings. No new data were collected. All audio recordings and derived feature tables were deidentified, with participants referred to only by study identifiers, and were stored on access-restricted, institutionally managed servers. Participant compensation for the parent trial is described in the original trial report [14].

Interview audio recordings were processed using a Python workflow designed (1) to standardize audio format and sampling characteristics; (2) to identify speakers and infer speaker roles (participant vs interviewer); (3) to generate time-aligned transcripts; and (4) to derive acoustic, prosodic, linguistic, sentiment, topic, and dyadic synchrony measures. The pipeline produces both segment-level datasets (eg, diarized turns, aligned words, and per-segment features) and participant-level summary tables.

Raw interview recordings were provided in common formats (eg, .mp3, .wav, and .wma). For each recording, study identifier, participant identifier, and (when available) week or session number were parsed from the filename using pattern-based rules. When week or session information was absent, a default value was assigned. These metadata were carried through all derived tables.

All recordings were converted to a mono, 16-kHz waveform to ensure consistent downstream feature extraction. Conversion was performed with FFmpeg when available. If FFmpeg failed or produced an invalid output, decoding and resampling were performed using “librosa.” Converted waveforms were cached in a derived folder and reused on subsequent runs when the cached file was nonempty and at least as recent as the source audio.

To facilitate quality control, file-level signal metrics were computed from the 16-kHz waveform. Framewise root-mean-square (RMS) energy was estimated using 2048-sample frames and a 256-sample hop. From the RMS distribution, we extracted a noise-floor proxy (10th percentile), median RMS, and an upper-energy estimate (95th percentile), and expressed these in linear units and decibels (dB), computed as 20·log₁₀(RMS+ε). Clipping was quantified as the fraction of samples with absolute amplitude ≥0.999. Total recording duration (seconds) was also recorded.

Speaker diarization was conducted using the WhisperX diarization interface (a wrapper around a PyAnnote-based pipeline), yielding a sequence of speaker-labeled time segments (start and end times in seconds). Because diarization labels are anonymous, we inferred speaker roles (participant vs interviewer) using a heuristic applied within each file.

Role inference used per-speaker diarization statistics, including total speaking time, number of segments, the proportion of short utterances (≤1.5 seconds), and time of first speech. The interviewer was preferentially identified as the speaker who (1) tended to speak earlier, (2) produced relatively short prompts more frequently, and (3) exhibited higher question use. The participant was identified as the remaining speaker with the greatest speaking time and longer average segment duration. Speakers not meeting the participant or interviewer criteria were labeled as “other,” and segments without an assigned role were retained with an “unknown” role label.

Automatic speech recognition (ASR) was performed using WhisperX. Each 16-kHz file was transcribed into text segments, and word-level timestamps were obtained via WhisperX alignment when enabled. Alignment used a wav2vec2-based alignment model (default: “facebook/wav2vec2-base-960h”) on a central processing unit and was executed in fixed-size batches of transcript segments to reduce memory demands and stabilize long-running jobs. If alignment failed for a file, the pipeline fell back to unaligned ASR segments.

Aligned word outputs were organized into a word-level table containing word text and onset and offset times. These word timestamps served as the basis for per-segment transcript reconstruction and timing-derived measures (eg, speech rate and pause counts).

To improve interviewer identification, question-mark counts were estimated from ASR text segments and attributed to diarized speakers by overlap: each ASR segment containing “?” was assigned to the diarized speaker whose segment window had the greatest temporal overlap with the ASR segment window, and question marks were accumulated by speaker. These per-speaker question counts were then incorporated into the role-inference heuristic to refine participant or interviewer assignments.

Diarized segments served as the primary unit of analysis (“turns”). For each diarized segment, aligned words were assigned to the segment when their start and end times fell within the diarized boundaries. Assigned words were concatenated to create a segment transcript and to compute timing measures.

Speech rate was computed as words per minute (WPM=60×number of words÷segment duration in seconds). Pauses were indexed as the number of interword gaps exceeding 250 ms (based on successive aligned word onset times). Linguistic features were computed from tokenized text using a consistent tokenization rule (alphabetic tokens with apostrophes) and stopword filtering. That is, all transcripts were processed with the same tokenization procedure and settings, so every linguistic feature (eg, word counts, lexical diversity, and sentiment counts) is computed on a comparable definition of “token.” The same rules for splitting text into tokens (eg, handling punctuation, contractions, casing, and numbers) were applied uniformly across all samples, avoiding measurement differences caused by inconsistent preprocessing. Derived metrics included total token count, number of unique tokens, lexical diversity (number of unique tokens divided by total tokens), hapax count and ratio (tokens occurring once), average word length, and the proportion or count of content words (nonstopwords). Normalized text (lowercased, tokenized, and stopwords removed) was retained for downstream lexical similarity computations.

To estimate the sentiment features, valence was computed from each segment transcript using a rule-based sentiment model, Valence Aware Dictionary and Sentiment Reasoner. Segment-level outputs included positive, negative, neutral, and compound sentiment scores. For speaker-level summaries, segment scores were aggregated using word counts as weights to reduce the influence of very short segments.

Acoustic features were extracted using openSMILE with the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPSv02 version 2.0) feature set at the “functionals” level, producing file-level summary descriptors. In addition, speaker-specific acoustic summaries were computed by concatenating all diarized segments for each speaker role (participant and interviewer) within a file and re-extracting the same eGeMAPS functionals from the resulting role-specific waveform. Acoustic summary statistics included pitch, loudness, spectral, and voice-quality measures summarized using standard statistics such as means, variability, and percentiles. Speaker-level acoustic tables also included total speaking time (sum of diarized segment durations) and the number of segments contributing to the summary.

Prosodic features were computed for each diarized segment from the 16-kHz waveform. Fundamental frequency (F0) was estimated using “librosa.pyin” (65-400 Hz) with 2048-sample frames and a 256-sample hop. For each segment, we summarized F0 mean, median, SD, minimum, and maximum across voiced frames and computed voiced fraction (proportion of voiced frames). Intensity-related features were summarized from the RMS envelope (mean, SD, minimum, and maximum), and spectral-temporal activity was indexed using the mean zero-crossing rate. Segment-level prosody was aggregated to speaker-level summaries within file and role.

To support downstream filtering and sensitivity analyses, segment-level text or timing features were merged with segment-level prosody and file-level quality metrics. Segment-level signal-to-noise ratio (SNR; dB) was computed as follows:

Additional derived indicators included segment RMS in decibels and a binary marker indicating whether the segment contained any aligned words. Summary quality indices were computed by speaker role, including mean or median SNR, proportion of segments with low SNR (<5 dB), proportion with no aligned text, mean WPM, and mean clipping fraction.

Topic modeling was conducted on participant segments meeting a minimum token threshold (default ≥5 tokens). Sentence embeddings were computed using a transformer encoder (default: “all-MiniLM-L6-v2”) and clustered with k-means (default k=12; reduced automatically when fewer segments were available). For interpretability, topic descriptors were computed using term frequency-inverse document frequency to identify top terms per topic, and each segment was assigned a topic label along with its similarity to the assigned centroid. Participant-level topic features were represented as topic-use proportions (topic counts normalized within participant and file).

Two complementary synchrony indices were computed to characterize coordination between participant and interviewer. First, a turn-pair synchrony dataset was created by pairing each participant segment with the temporally nearest interviewer segment (nearest center time). For each paired turn, we computed absolute differences in timing (center-time distance), prosody (difference in mean F0 and mean RMS), speech rate (WPM difference), affect (difference in compound sentiment), and lexical alignment (Jaccard similarity over normalized token sets). Second, prosodic synchrony was summarized in fixed temporal bins (default 1-second bins) by averaging participant and interviewer features within each bin and computing within-file correlations and mean absolute differences between roles for F0, RMS, and sentiment. These synchrony summaries were exported at the participant-file level.

The pipeline exported intermediate and derived datasets as both Parquet and CSV files, including diarization segments, aligned words, segment-level text metrics, acoustic and prosodic features, sentiment summaries, topic outputs, quality-control variables, synchrony metrics, and a consolidated participant-level feature table. Participant-specific copies of each dataset were additionally written to participant folders to support participant-level analyses and auditing.

For statistical analyses, we constructed a consolidated, analysis-ready feature matrix by merging participant-level summaries across domains using participant ID and recording filename as keys. This matrix comprised 161 participant-level derived features: acoustic descriptors (n=90), prosody (n=13), linguistic content and fluency measures (n=10), sentiment (n=6), quality-control metrics (n=17; 9 segment-level summary indices and 8 file-level signal-quality metrics), topic composition (n=12 topic-proportion features), and dyadic synchrony (n=13; 6 prosodic synchrony indices and 7 turn-level synchrony summary indices). As described in the Statistical Analysis section, the acoustic descriptors were subsequently summarized by domain composites, and the feature set was screened and reduced before analysis. The full list of features entered into the analyses is provided in Table S7 in Multimedia Appendix 2.

Analyses were conducted at the participant level using baseline (week 0) and follow-up (week 4) symptom outcomes, the MASQ-AD and QIDS-SR. Although the focus of this report is anhedonia, we included the QIDS-SR for comparison. The feature matrix comprised acoustic, prosodic, linguistic, sentiment, topic, synchrony, and ASR quality measures. Data quality was summarized for each participant (low SNR in >30% of files, empty transcripts or ASR failures in >50% of files, or mean transcript length <3 words). All participants were retained to preserve power, and analyses are interpreted with this data-quality caveat in mind. In sensitivity analyses of the cross-sectional anhedonia associations, we (1) re-estimated the associations after excluding participants flagged for potentially problematic data quality and (2) computed partial correlations adjusting for total speech quantity (total tokens); these sensitivity analyses used parametric correlation tests (Tables S1 and S2 in Multimedia Appendix 2).

Feature preprocessing included removal of predictors with >20% missingness or zero or near-zero variance, followed by median imputation of remaining missing values. Because the acoustic feature set was high-dimensional relative to the sample size, the acoustic domain was represented by theory-driven domain composites constructed for each acoustic subfamily (pitch, jitter, shimmer, mel-frequency cepstral coefficients [MFCCs], formant, loudness, spectral, voice quality, and prosody) by averaging z-scored features within each domain. These composites replaced the individual acoustic variables in the predictor set. Because several segment-level prosodic and temporal features were also retained as individual predictors, the prosody composite overlaps with those features. The domain composites are therefore intended as coarse summaries of feature families rather than mutually independent predictors. After this reduction, 77 features were entered into the analyses for each outcome (68 nonacoustic features and 9 acoustic domain composites).

Primary inference relied on univariate Pearson correlations between each feature and each outcome, with statistical significance assessed using permutation tests (1000 label permutations). For week-4 outcomes, the week-4 symptom score was residualized on baseline symptom level and ABM assignment, and each voice feature was correlated with these residuals. These week-4 associations therefore reflect where a participant stood at week 4 relative to the level predicted by their baseline severity and treatment assignment, rather than an individual change in symptoms over time. Significance for the adjusted associations was evaluated with the same 1000-permutation procedure. We summarized the strongest associations for each outcome and used volcano plots for visualization. Analyses were performed in R (R Foundation for Statistical Computing), with fixed random seeds to ensure reproducibility.

Because these analyses are explicitly exploratory and hypothesis-generating, we report unadjusted P values and emphasize effect sizes and consistency across outcomes. To limit the multiple-testing burden, we took 2 complementary steps: we reduced the number of tests by representing the acoustic domain with 9 composites, and we adopted a conservative interpretive threshold of P=.01 for identifying associations of interest. We note the exact number of tests: each outcome was tested against 77 features, and the 4 primary analyses reported here (MASQ-AD and QIDS-SR, each examined cross-sectionally at baseline and in baseline-adjusted week-4 models) comprise 308 univariate correlation tests. We did not apply a formal familywise or false-discovery correction, which would be overly conservative for an exploratory screen of this kind. Permutation P values provide an empirical null tailored to the data and help calibrate significance under minimal distributional assumptions, but they do not themselves correct for multiple testing. We did not fit multivariate models because the high feature-to-sample ratio risks unstable estimates and overfitting, so we prioritized transparent univariate screening as a first-pass exploratory step. All findings are interpreted as preliminary signals that require independent replication.

In this study, we conducted an exploratory analysis of whether speech-derived features from a semistructured clinical interview are associated with individual differences in anhedonia severity in a sample of adults with elevated depression, and whether baseline features are associated with anhedonia 4 weeks later after adjusting for baseline severity. Using a multimodal framework spanning acoustic, prosodic, linguistic, and topic-based features, and applying a conservative P=.01 interpretive threshold, we observed a coherent cross-sectional profile associated with higher baseline anhedonia: reduced speech output (fewer content and total words and shorter speaking duration) alongside higher lexical-diversity ratios. Weaker, nonsignificant associations pointed toward reduced activity-related (work or school) content, more negative sentiment, and attenuated pitch variability.

In baseline-adjusted week-4 analyses, no baseline feature was associated with week-4 anhedonia at the P=.01 threshold, although the strongest nonsignificant effects again involved reduced speech output and engagement. The only week-4 association to exceed the significance threshold was for general depression severity, where lower baseline voice perturbation (jitter) predicted higher-than-expected week-4 symptoms.

Together, these preliminary findings are consistent with the possibility that anhedonia is expressed in speech more through reduced behavioral engagement and altered linguistic efficiency than through uniform flattening of vocal expressivity, and they are compatible with a dimensional approach to speech-based phenotyping of depressive symptoms. Because these associations were estimated in a small sample across many features, they should be regarded as hypothesis-generating rather than confirmatory even under the conservative threshold applied here.

One notable finding was that higher anhedonia was associated with lexical brevity and diversity (shorter speech, higher hapax ratio and lexical diversity, and fewer content words). We interpret the diversity finding cautiously, because hapax ratio and lexical diversity (type-token ratio) are mechanically sensitive to text length. With fewer total words, the proportion of unique words tends to be higher simply because there are fewer opportunities to repeat, so apparently “greater diversity” may be largely an artifact of reduced speech quantity rather than richer content. In a semistructured interview, short responses to multiple prompts can introduce new vocabulary without much repetition, further inflating diversity ratios. The concurrent finding of fewer content words is consistent with this account. We therefore treat the lexical-diversity associations as potentially artifactual and mechanically driven by speech quantity rather than as evidence of more elaborate or information-dense speech [17]. Consistent with this, a sensitivity analysis adjusting for total speech quantity substantially attenuated the lexical-diversity associations (Table S2 in Multimedia Appendix 2).

Reduced speech output was particularly evident when participants responded to the HRSD-17 interview item assessing recent activity, a domain closely tied to behavioral engagement in depression. Topic-model features pointed in the same direction, though this association was nonsignificant at our P=.01 criterion: reduced work- or school-related content at baseline (r=–0.31; P=.04) and, in the baseline-adjusted week-4 analyses, less check-in-related content among participants with higher-than-predicted week-4 anhedonia (r=–0.31; P=.06). These patterns are consistent with contemporary models of anhedonia in depression that emphasize impairments in motivation, effort allocation, and goal-directed behavior, often described as motivational or anticipatory anhedonia, rather than a simple reduction in hedonic capacity [4]. Reduced verbal output and narrower topical focus during activity-oriented clinical interviews may therefore index clinically meaningful aspects of behavioral activation and reward pursuit [18], although our cross-sectional design cannot establish that these speech features are caused by these processes.

In addition to less speech, anhedonia was associated with less expressive speech. Affective flattening is often described as reduced variability in vocal affect, with less modulation of pitch, intensity, and prosodic contours [17,19,20]. The observed flatter F0 variability, although nonsignificant at the P=.01 criterion, is directionally consistent with this account, suggesting a narrower expressive range, and the slightly higher pitch floor could indicate a constrained pitch band anchored higher but with less dynamic movement. Stratified analyses in Table S3 in Multimedia Appendix 2 indicated that the pitch associations were broadly consistent across sex. However, given that only 9 participants were male, these within-sex analyses are themselves underpowered and should not be taken as strong evidence that the findings are sex-invariant. Together, these prosodic features offer convergent but tentative evidence that anhedonia may be reflected not only in what participants say but in how vocally expressive their speech sounds.

The baseline-adjusted week-4 associations were weaker than the cross-sectional findings and based on a much smaller subsample. For anhedonia, no baseline feature was associated with week-4 symptoms at the P=.01 threshold. For general depression severity, one association exceeded the threshold: lower baseline voice perturbation (the jitter domain composite) was associated with higher-than-predicted week-4 severity. Notably, across analyses, anhedonia was more consistently associated with temporal and quantity-based measures than with pitch-based indices, which are known to vary across tasks and samples [21,22].

These findings are consistent with the broader argument for targeting specific symptom dimensions rather than global depression severity when developing speech-based models. Depression is a heterogeneous syndrome comprising partially dissociable symptom clusters that may differ in their behavioral and neurobiological expression [23-25]. In the current study, speech features showed somewhat stronger and more interpretable associations with anhedonia than with overall depression severity, which would be consistent with speech more directly indexing motivational and behavioral engagement processes central to anhedonia. We note, however, that this comparison is not definitive: it could partly reflect restricted range in total depression scores, given that study entry required elevated depression (Figure S1 in Multimedia Appendix 2), as well as measurement differences between the self-reported MASQ-AD and QIDS-SR. If replicated, speech-derived markers of verbal engagement and topic content could eventually complement symptom scales as low-burden, ecologically valid indices of motivational engagement. This clinical potential remains aspirational at the current stage of evidence. Future studies should assess additional depressive dimensions (eg, psychomotor retardation, cognitive control, rumination, and anxiety-related arousal) to determine whether distinct speech signatures map onto different facets of depression [26].

Several limitations temper these conclusions. First, and most importantly, the sample was small (analytic: N=45; cross-sectional: n=44 for anhedonia and n=45 for depression severity; longitudinal: n=36). With this many derived features, small samples make individual associations vulnerable to instability and to both false positives and false negatives, and they preclude reliable multivariate or machine learning models. The longitudinal analyses in particular were powered to detect only large effects and rest on at most a single significant association, so they are best read as hypothesis-generating.

Second, we tested many features, and although we reduced the feature set and adopted a conservative P=.01 threshold, we did not apply a formal correction for multiple comparisons. Each outcome was correlated with 77 features, and the 4 reported analyses comprise 308 univariate tests. Although permutation testing provides a data-driven null, it does not by itself control the familywise or false-discovery rate. The associations should therefore be interpreted as preliminary and exploratory rather than confirmed effects.

Third, the sample was predominantly female (36/45, 80%) and was drawn entirely from a depression intervention trial. The sex imbalance is a substantive limitation because several vocal features, particularly pitch, are sex-sensitive. Our within-sex analyses were underpowered and cannot rule out that some associations differ by sex. Because all participants had elevated depression and the speech came from a depression interview, we also cannot determine whether the observed features are specific to anhedonia versus depression more broadly, and the restricted range in depression severity limits comparisons across symptom dimensions. These features of the sample limit generalizability to community, nonclinical, or transdiagnostic populations, and the claims in this report should be read as applying to adults with depression assessed in a clinical interview rather than to speakers in general.

Fourth, data quality was variable. Of the 45 analytic participants, 14 were flagged as having potentially problematic data (low SNR, frequent ASR failures, or very short transcripts); yet, all were retained to preserve power. Retaining lower-quality recordings adds measurement noise and further increases uncertainty about the stability of the reported associations. In a sensitivity analysis excluding the flagged participants, the lexical-diversity associations were essentially unchanged, whereas the speech-quantity associations attenuated but retained the same direction (Table S1 in Multimedia Appendix 2). Results should nonetheless be interpreted with this in mind, and future work should preregister quality thresholds.

Finally, the speech data were derived from a semistructured clinical interview, which may constrain the range and expression of vocal and linguistic features relative to more naturalistic contexts. Interview prompts can shape speech quantity, topic distribution, and prosody in ways that may not generalize to everyday communication, so the observed associations may reflect how depressive symptoms manifest under evaluative, goal-directed conversation rather than spontaneous speech. Future studies incorporating more ecologically valid speech samples, such as passive audio capture or ambulatory voice diaries, will be important for assessing generalizability.

Within the constraints of a small, exploratory, predominantly female clinical sample, this study adds to an emerging literature, suggesting that anhedonia may be detectable in the speech of adults with depression. Higher anhedonia was associated most robustly with briefer speech (fewer total and content words and shorter duration), with weaker, nonsignificant associations with more negative sentiment, flatter pitch variation, and less work or school topic content, a pattern more consistent with reduced behavioral and verbal engagement than with uniform flattening of expressivity. The accompanying increase in lexical diversity is most plausibly an artifact of reduced speech quantity. More broadly, the results illustrate both the promise and the current limits of speech-based phenotyping of specific symptom dimensions: if these signals replicate in larger, more diverse, and more naturalistic samples with appropriate multiplicity control, voice-derived indices of verbal engagement could become useful, low-burden complements to existing symptom measures and may also provide an important insight into how anhedonia manifests in natural behavior.