The identification and diagnosis of depression through clinical interviews are often slow and unreliable (1–4). About half of cases go unrecognized: in a meta-analysis of 41 studies, recognition accuracy of depression by general practitioners was 47.3% (5). Therefore, accurate and fast ways to identify cases of depression will have major clinical benefits.

Novel applications of computational methods are making some inroads into this problem. For example, a review of 14 studies indicated that both sensitivity and specificity of diagnostic performance of deep learning models were higher than that of health-care professionals (6). In the last decade, there have been interests in the ability of exploiting vocal biomarkers to identify depression with machine learning methods to investigate whether voice can be used as an auxiliary tool to assist clinical diagnosis (7–11). Previous research mainly focused on examining the ability of vocal features in classifying individuals with depression and healthy population (12–17), and F-measure of relevant classifiers reached 0.9 (12). These findings suggest that vocal biomarkers may have discriminatory potential in identifying depression. While the differential diagnosis is complicated by the presence of other mental disorders.

The prevalence of mental disorders, according to a study in China, showed that the weighted prevalence of any disorder (excluding dementia) was 16.6% (95% CI 13.0–20.2) during the 12 months before the interview (18). Clinically, the psychiatric diagnosis necessitates distinguishing depressed individuals not only from the healthy ones, but also from other mental illnesses with similar mood symptoms or similar voice patterns. It needs to be stressed that voice conveys emotion related information. In the field of affective computing, voice features were shown their abilities in recognizing different kinds of emotions (19, 20). Both bipolar disorder (BD) and schizophrenia exhibit symptoms comparable to depression. According to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-4) (21), BD involves both depressive and manic episodes, while schizophrenia patients with negative symptoms report anhedonia (loss of ability to experience pleasure). The specificity of alterations in vocal characteristics to depression has not been rigorously investigated.

Variations in vocal features have been observed in schizophrenia and BD patients. Acoustic analyses of speech in schizophrenia have revealed subtle aberrancies in pitch variability associated with flat affect, as well as more pronounced deviations in properties such as percentage of speaking time, speech rate, and pause duration that correlate with symptoms of alogia and blunted affect (7). Espinola and her colleagues (22) constructed classification models based on speech samples from 31 individuals (20 of whom had a prior diagnosis of schizophrenia and 11 healthy controls). The classifiers attained an accuracy of 91.76% in distinguishing between groups. Regarding BD, studies have utilized vocal features to predict patients’ emotional states (e.g., depressive, manic, mixed states). Classification analyses yielded an area under curve (AUC) of 0.89 (9, 10). Another study investigated whether vocal features acquired via verbal fluency tasks could accurately differentiate mixed states in BD using machine learning methods. And results showed that for depressive versus mixed depressive episodes, the F-measure was 0.86, while for hypomanic versus mixed hypomanic episodes, the F-measure was 0.75 (23). These studies showed that voice features may also be informative for other psychiatric diagnosis. It should be noted that several studies have examined the utility of vocal features in developing classifiers for several mental disorders. A study (24) employed polytomous logistic regression analysis of vocal features to discriminate among healthy controls (n = 23), individuals with bipolar disorder (n = 8), and those with major depressive disorder (n = 14). The model attained 90.79% accuracy in classifying participants into the three diagnostic groups. Another study (25) proposed a methodology to support the diagnosis of several mental disorders using vocal acoustic analysis and machine learning. The results showed that random forests with 300 trees achieved the best classification performance (75.27% for accuracy) for the simultaneous detection of major depressive disorder (26), schizophrenia (20), BD (15), generalized anxiety disorder (4), and healthy controls (12). However, the datasets of above two studies were imbalanced for each group. The imbalanced dataset problems become more complicated in multi-class imbalanced classification tasks, in which there may be multiple minority and majority classes that cause skewed data distribution. And machine learning algorithms tend to favor the majority class samples, hence damaging the multi-classification results (27, 28). Moreover, extensive comparisons between mental disorders and healthy controls may offer more information about the effectiveness of voice for clinically complex differential diagnosis.

Various speech features are indicative of depression. Mel-frequency cepstrum coefficients (MFCCs) constitute the most prevalent vocal features employed in speech recognition systems and psychiatric condition classification models (26). MFCCs are obtained by extracting frequency spectral features of the speech signal using the short-time power spectrum, mapping these features onto the Mel scale to better present auditory characteristics, and then obtaining MFCC coefficients through cepstrum analysis that can characterize the speech envelope (12, 29). Multiple studies have demonstrated the utility of MFCCs in developing classification models for depression detection (30–32). For example, Di et al. (29) employed MFCCs to classify major depression patients and healthy individuals, area under curve (AUC) reached 0.8.

The identity vector (i-vector) approach, grounded within the total variability framework, represents the state-of-the-art technique for speaker verification (12, 29, 32, 33). The total variability framework offers an effective means of capturing speaker- and channel-related variability in a low dimensional subspace. i-vectors are highly informative for encoding cepstral variability. Classification models based on i-vectors demonstrated capacity for identifying depression with high accuracy. For instance, prior work found i-vector based model outperformed a baseline model defined by KL-means supervectors (32). Nasir and his colleagues (33) used i-vectors to investigate various audio and visual features for classification, reporting high accuracy with i-vector modeling of MFCC features. Indeed, one study demonstrated a 40% improvement in predictive accuracy (F-score) with the i-vector methodology (12). And Di et al. (29) observed a 14% enhancement in model performance (AUC) with i-vectors relative to MFCCs alone. Although the participant cohorts were exclusively female in both studies, the results demonstrated the promise of i-vectors for enhancing the accuracy of machine learning models for depression classification.

The objective of this study was to evaluate the efficacy of vocal features as differential diagnostic markers for depression compared to other psychiatric disorders. Three binary classification paradigms were employed in total: (1) the capacity of voice features to distinguish any psychiatric condition (depression, bipolar disorder, schizophrenia) versus healthy controls at baseline; (2) the ability of voice features to differentiate a specific psychiatric illness from healthy controls; (3) the capability of vocal features to distinguish between discrete psychiatric disorders in a pairwise manner. Among these paradigms, the first one was served as a baseline to determine whether the dataset achieved performance commensurate with existing research, as well as a benchmark for model performance under other framework conditions. The second and third paradigms were employed to systematically evaluate the capacity of vocal characteristics to distinguish between case and control groups.

In this research, we investigated the ability of vocal biomarkers to classify various health conditions. Following matching, descriptive statistics showed no differences in demographic variables between the case and control groups for any task, addressing potential threats to validity (15). Logistic regression models based on MFCC-derived i-vectors were developed for all classification tasks. Results indicated vocal features may assist differential diagnosis of depression, albeit with varying degrees of effectiveness across classification tasks.

The AH model examined the baseline ability of vocal features to distinguish clinical from healthy groups. Classifying the healthy group and the any-disorder group yielded an F-score of 0.82 and AUC of 0.79, demonstrating vocal features can distinguish mental disorders from the health.

Furthermore, the DH, BH, and SH models investigated the ability of voice features in distinguishing specific disorders from health. For DH, BH, and SH tasks, F-scores ranged from 0.73 to 0.80, AUC scores from 0.75 to 0.80. Model comparisons showed there were no significant differences in the pairwise comparisons among the AH and the three mental illnesses versus healthy control classification tasks. Our results about DH are consistent with the existing findings (8, 12, 15, 47–50). For example, a previous study examined the significance of the association between voice features and depression using binary logistic regression, and the actual classification ability of voice features on depression using machine learning method, results showed that the contribution effect of voice features reached 35.65% (Nagelkerke’s R²), further classification model achieved 81% of F-measure (15). We set this classification model as another baseline for the reference of further classification. The results of BH and SH classification models are also consistent with existing studies. For instance, one study investigated whether voice features from naturalistic phone calls could discriminate between BD and healthy control individuals, results showed that compared to the control group, BD was classified with a 0.79 sensitivity and 0.76 AUC (51). For schizophrenia detection, Tahir et al. (52) classified schizophrenia patients and healthy controls with multilayer perceptron and the accuracy achieved 81.3%. It was also suggested that speech abnormalities, related to flat affect and alogia, have been a hallmark feature of schizophrenia, and are often associated with core negative symptoms and social impairment (7). Our results indicate that MFCCs capture information distinguishing depression, BD and schizophrenia from health.

The DB, DS, and BS models further evaluated the performances of voice features on pairwise classifications among the three mental disorders. DS (F-score = 0.83; AUC = 0.83) and BS (F-score = 0.91; AUC = 0.92) models showed promise. The BS model had the highest F-score and AUC across all seven models. However, the DB model performed the worst (F-score = 0.44), with an AUC score of 0.50 indicating voice features barely distinguished depression and BD. Further pairwise model comparisons showed no significant differences among the AH, DH, BH, SH, DS, and BS models. However, DB model performance was significantly worse than that of BS. Results indicate that voice features could help distinguish depression from disorders with similar mood symptoms.

Differing symptoms and vocal characteristics likely explain the results of differential diagnosis. For example, depression is associated with monotonous, hoarse, breathy or slurred speech reflecting anhedonia and sadness (53–58). Schizophrenia is linked to poverty of speech, increased pauses, distinctive tone/intensity associated with core negative symptoms like flat affect, decreased emotional expression and difficulty controlling speech to express emotion properly (59–65). Given that Espinola et al. (25) found vocal features distinguishing major depressive disorder, schizophrenia and other disorders, our MFCC derived i-vector approach may reveal subtle differences in anhedonic/sad depression symptoms versus alogia/flat affect in schizophrenia. However, further voice difference investigation between depression and schizophrenia is still needed to offer more explanation.

Another possible reason is that the i-vectors catch disorder-relevant information. To extract i-vectors, first the method learns shared case/control information, then removes shared components, yielding i-vectors capturing key individual differences (29, 43). Here, i-vectors captured distinct vocal information for different mental disorders.

Overall, vocal features could provide clinical value in distinguishing depression from selected disorders, and model improvement is helpful, especially for closely related conditions. This analysis establishes a foundation for future studies exploring vocal biomarkers for differential diagnosis of psychiatric disorders.

This research was unable to clearly differentiate between depression and BD. There appears to be substantial symptom overlap between these two conditions. Furthermore, the oscillation between manic and depressive states in BD can be subtle, as demonstrated by Grande et al. (66). Researchers have posited that there might be continued presence of subsyndromal residual symptoms during recovery from major affective episodes in bipolar disorder, and residual symptoms after resolution of a major affective episode indicate that the individual is at significant risk for a rapid relapse and/or recurrence, augmenting the intricacy of symptom presentation even during euthymia (67). Depressive symptoms may have been present in euthymic bipolar patients in this study, confounding diagnostic classification. This observation highlights the complexity of the euthymic bipolar condition. Future research should examine symptom and voice differences between BD phases and depression. A nuanced understanding of markers that distinguish unipolar depression from BD could sharpen diagnostic precision.

This study has limitations. We examined only depression, BD and schizophrenia in a small cultural sample, limiting generalizability. It is worth noting that this study lacks an anxiety disorder group, which also exhibits affective symptoms similar to depression. As the data came from a previous project in China, anxiety diagnoses were not included. Future work will gather systematic data on symptoms, clinical phase, psychological factors like emotion, cognition, and severity in these and other disorders, such as anxiety. This could yield insights into voice differences between depression and other conditions, enhancing the clinical value of vocal biomarkers.

This research systematically explored the ability of vocal biomarkers to distinguish depression from disorders with similar affective symptoms. Findings suggest vocal features could aid differential diagnosis for depression in clinical practice. Future research should investigate mechanisms by which vocal features differentiate depression and other disorders, and develop more advanced machine learning models so voice can enhance clinical diagnosis.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Institutional Review Board (IRB) of Institute of Psychology, Chinese Academy of Sciences. The patients/participants provided their written informed consent to participate in this study.

WP was in charge of overall research, data analysis, and drafting the paper. FD was in charge of offering professional guidance and advice for mental illnesses as a clinical specialist. XW helped on data analysis. BH and WZ helped with the analysis and revising the manuscript. TZ was in charge of collecting and organizing the data. All authors contributed to the article and approved the submitted version.

This work was supported by the Fundamental Research Funds for the Central Universities (CCNU21XJ021), Knowledge Innovation Program of Wuhan-Shuguang Project (2022020801020288), and the Research Program Funds of the Collaborative Innovation Center of Assessment toward Basic Education Quality (2022-04-030-BZPK01).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.