ABSTRACT

Background

Clinicians aspire to predict the emergence of Bipolar Disorder (BD) in a timely manner. To accomplish this, markers reflecting mental states that can be gathered non‐invasively and at large scale are needed. Here, we systematically evaluate evidence relating speech‐based markers to mood states in BD.

Methods

We searched Medline and Google Scholar for all published studies in English up to February 2026 on the use of speech markers in BD. We undertook thematic analysis on abstracts using topic modeling and a qualitative gap analysis to identify potential opportunities for future research.

Results

43 out of 867 studies were included after screening. Topic analysis revealed an emerging focus on mapping mood states to automated speech features. Most studies focused on cross‐sectional detection of bipolar mood states, or BD as a diagnosis, rather than the prediction of upcoming mood states. Speech features distinguished BD from schizophrenia, depression, and healthy controls. Manic states were characterized by quantifiable measures of pressured speech, derailment, grammatical errors, and word repetition; depressive states by an increased use of personal pronouns, reduced verbal fluency, and speech quantity. Overall, attempts to replicate observations were limited.

Conclusion

Acoustic and lexical‐semantic markers vary with manic, psychotic, or depressive states. At present, the evidence is insufficient for clinical utility in relapse prediction, response monitoring, or diagnosing mixed episodes or state changes in BD. We recommend that future research leverages the growing capabilities of natural language processing through longitudinal and cross‐linguistic studies to strengthen the evidence base and advance the clinical utility of speech markers for BD.

1 Introduction

Early detection of impending states in bipolar disorder (BD) can aid in earlier and more accurate intervention, as its presentations—whether states with manic, hypomanic, or depressive features—often mimic phenomenologically related disorders (unipolar depression, non‐affective psychosis, or substance‐induced episodes). The fluctuations in emotions, energy, and cognitive functions associated with BD are often unpredictable and, in some cases, occur repeatedly despite treatment. These repeated recurrences (relapses) are associated with a higher risk of suicide [1, 2], especially in depressed or episodes with mixed features [3]. Thus, reliable prediction of the emerging mood states when they are subtle may allow early detection of both onset and recurrences and thus timely interventions in BD.

While the clinical gains achieved by blood‐ and brain‐based biomarkers have been limited [4, 5], two other objective measures have gathered pace in recent years. The first is digital phenotyping using wearables that detect activity patterns, extensively reviewed elsewhere [6, 7]. The second is the use of speech recordings to detect changes in voice (acoustics) and communication patterns (i.e., the form and content of language). Clinicians routinely use observed speech (i.e., sampling of interpersonal communication) as a critical source of information to rate and measure symptoms. Speech can be economically and non‐invasively obtained and has been hailed as a biosocial marker for mental illnesses [8]. We can now objectively quantify clinically pertinent features from speech using automated Natural Language Processing (NLP) approaches. Here, we review the work done to date in BD using various speech‐based markers.

Readily accessible and objective markers of the illness have three key applications in BD. Firstly, in differentiating first episode BD and Major Depressive Disorder (MDD) from non‐affective psychosis. Confident identification on the basis of cross‐sectional assessment of symptoms [2, 9] or cognitive measures [10, 11] alone continues to be a challenge but has important therapeutic and prognostic implications. Secondly, in predicting the emergence of mixed features after remission from a depressive episode. Hypomania is often the first sign of BD in 12%–17% of patients being treated for depression, but is often under‐reported by patients and missed by clinicians [12, 13, 14], increasing the risk for further antidepressant‐related mood destabilization [15, 16]. Third, in identifying episodes with mixed features. The presence of features with opposite polarity during a mood state is increasingly recognized to be more common (states with mixed features [17]) but is often missed in practice, leading to inadequate treatment [18]. Developing readily accessible markers can improve the diagnostic accuracy (psychosis vs. mania), specify clinical states with precision (mixed features, upcoming relapses), and inform us of the likely disease trajectory (unipolar to bipolar transitions, or a misdiagnosis of BD depressive episodes); all of these are well‐known unmet needs in the clinical care of BD [19]. In addition, potential “trait” markers may identify individuals at risk of first episode mania, allowing for early intervention that may have superior outcomes [20, 21].

Language disturbances have been one of the core diagnostic symptoms of both manic and depressive states. As per the DSM‐5, speech in mania is “rapid, pressured, loud and difficult to interrupt…(patients) may talk continuously and without regard for others' wishes, often in an intrusive manner or without concerns for the relevance… sometimes characterized by jokes, puns, amusing irrelevancies, and theatricality” [22]. In depression, speech changes are captured as psychomotor retardation: “slowed speech, thinking, increased pauses before answering; speech that is decreased in volume, inflection, amount, or variety of content, or muteness” [22]. Thus, speech features constitute the diagnostic criteria for mood states, unlike many other digital phenotype markers (e.g., geolocations, step counts, call and text logs, electrodermal hypoactivity), which are only indirectly related to the constructs of depression and mania [23, 24]. Several cognitive measures relevant to BD (e.g., verbal fluency) [25] can also be remotely obtained using digital tools [26, 27], but speech can be sampled in completely naturalistic settings (e.g., during a routine clinical interaction, a social conversation, etc.) and repeatedly acquired for longitudinal studies without invasive or intrusive procedures. This unique combination of construct and ecological validity makes speech markers well positioned for clinical applications in BD.

Natural Language Processing (NLP), a branch of artificial intelligence (AI), combines conventional computational linguistics with speech technologies (e.g., automatic speech recognition) in an attempt to make human language machine usable. By using the inherent hierarchical statistical structure of human language, voice‐related information in speech, and contextual features of discourse, NLP tools allow us to extract acoustic or paralinguistic, syntactic, semantic, sentimental, and pragmatic markers from speech recordings [28, 29, 30]. In effect, clinical NLP can be used to analyze any spoken or written material, by patients or by caregivers. While a review has highlighted NLP's potential for characterizing BD when applied to written text (Electronic Health Records and Reddit/Twitter posts), here we specifically focus on speech markers [31].

We systematically review prior studies examining speech in BD using automated (NLP) approaches and synthesize the features of speech and language that are specific in identifying and predicting the various mood states in BD. Based on a long line of rating‐scale‐based assessments of speech, language, and thought disorders in BD [32, 33, 34], verbal fluency studies [25, 35, 36, 37], and first‐person reports [38], we expected to see (1) illness‐specific and (2) state‐specific speech and language variables in BD, and potentially (3) the use of speech markers to predict impending mood states.

2 Methods

2.1 Search Process

We undertook a systematic search of the literature (as per our pre‐registration [39]) using the Preferred Reporting Items for Systematic reviews and Meta‐Analyses (PRISMA) to identify all published human studies written in English pertaining to using speech markers in bipolar speech during any mood state (see Figure 1) [40]. An extension was applied upon request to include studies published until February 2026 which is a deviation from the pre‐registered protocol date of June 2024. We searched MEDLINE (via PubMed) and Google Scholar databases in February 2026. Search terms included a combination of keywords and Medical Subject Headings (MeSH) terms related to bipolar disorder and speech/language. For example, in PubMed we used the following combination: (“Natural Language Processing” [tiab] OR “Speech” OR “verbal” AND (Bipolar OR mania OR manic) NOT review). Variations of these terms were also used to ensure no relevant hits were missed. F.Z., J.A., and A.V. conducted the literature search and independently used Covidence (Veritas Health Innovation, Melbourne) to screen titles and abstracts against the inclusion criteria after removing duplicates. We used the following inclusion criteria: (1) Studies on human subjects with bipolar disorder and mean age above 18 compared to healthy or psychiatric control groups or within‐subject mood‐state comparisons; (2) Measured and analyzed spoken and/or transcribed speech; and (3) articles published in English. Full texts of relevant studies were assessed for eligibility. We then added further studies to the pool by screening the bibliography and hand‐searching all citations received by the identified studies via Google Scholar. A consensus among independent reviewers was required to determine the feasibility of including each study. Data extraction was performed by one reviewer and subsequently checked by another to ensure accuracy. Any conflicts during the selection or extraction process were resolved through a consensus meeting.

We excluded non‐empirical studies (case reports/case series). Additionally, studies focused on high‐risk subjects without a diagnosed bipolar disorder; studies reporting verbal outputs that were either restricted (e.g., scripted conversations) or likely to have been edited after production (e.g., written reports and social media texts), studies not testing language production but only comprehension or recall were not eligible. Studies with retraction notices were also excluded.

2.2 Topic Analysis

From the included papers, we extracted the text on introduction/background, methods, results, and discussion from abstracts. We created word clouds per section, to visually assess if the content aligned with the purpose of our review. We then applied thematic analysis to the abstracts using topic modeling with Non‐negative Matrix Factorization (NMF) to study the time‐varying semantic trends in this field. Preprocessing involved applying a Term Frequency‐Inverse Document Frequency (TF‐IDF) vectorizer limited to 1000 terms to retain only the most informative words while excluding common stopwords. The NMF model was then applied to the TF‐IDF matrix to extract latent topics, with the number of topics set to six based on an exploratory assessment of coherence and distinctiveness, balancing interpretability with adequate coverage of thematic variation in the abstracts. Each topic was represented by its top five keywords, reflecting the primary themes in the abstracts. To analyze temporal patterns in topic distribution, abstracts were grouped by publication year into 5‐year intervals (e.g., 2000–2004, 2005–2009) which ensured sufficient data density within each period.

2.3 Data Extraction

We extracted the following variables from the full text of each study: design (case–control, longitudinal); sample size (patients/controls); mean age of the samples, % of females in each patient cohort; mood states of subjects (euthymic/manic/depressive); mode of data acquisition (i.e., phone data, in‐person interviews, video calls); nature of speech tasks (e.g., picture description, social conversations, clinical interview, list generation); the software used for automated language analysis (e.g., openSMILE, PRAAT); major variables of interest (e.g., pause length, semantic similarity); machine learning algorithms used and their performance (if applicable). Additionally, the key variables for BD or BD state detection were extracted. Acoustic variables were grouped based on the review by Low et al. [41] into four categories: source, filter, spectral, and prosodic/melodic features. Semantic variables were grouped in seven categories: lexical, syntactic, coherence, content, pragmatic, affective, and speech graph features, a division based on key variables. The quality of included studies was assessed (see Table 1 for quality rubrics), with each study's total quality score recalculated to a score between 0% and 100%.

3 Results

3.1 Study Selection

Figure 1 shows the PRISMA flowchart of our systematic study search. From the initial screening of 867 studies, 43 were selected investigating speech in patients with BD. As shown in Figure 2, the topic analysis confirmed that the selected papers aligned with our primary focus, that is, analysis of speech features pertaining to detection or measurement of mood states in bipolar disorder or the mood states thereof.

A heatmap, also shown in Figure 2, illustrates the trend over the years, demonstrating a focus on thought, emotion, voice, syntax, speech features, and comparisons to schizophrenia. Of these, an increased focus on speech features in relation to mood is notable in recent years. Initial papers had a focus on syntactic or linguistic features such as verbs and pronouns, while papers since 2015 focused relatively more on acoustics and machine learning (with key terms such as “features,” “speech,” and “area under the curve” indicating classification algorithms).

3.2 Study Characteristics

Sample sizes of BD participants ranged from 4 [42] to 141 [43]. Of the 43 included studies, 14 had a longitudinal component (ranging between 2 days and 16 months of speech data collection); the rest were cross‐sectional. Some studies acquired multiple recordings but only analyzed cross‐sectional results [44, 45]. 18 studies compared BD to healthy controls, 13 with schizophrenia, and 12 with depression. With respect to the mood states captured in BD, 28 (65%) included manic or hypomanic, 21 (49%) included depressed, 27 (63%) included euthymic, and 9 (21%) included mixed states (as many included multiple states, the sum exceeds 100%). The quality of the identified studies varied notably between 20% and 100% in our scoring scheme, see Table 2.

To distinguish BD from other diagnoses and healthy controls, or different BD states from each other, 58% of studies (25/43) used classification algorithms (6 using lexico‐semantic information; 19 using acoustic features). The most common algorithms used were Support Vector Machine (6 studies) and Random Forest (8 studies). Only two studies investigated subclinical changes in mental states [46]; however, none predicted upcoming state changes.

3.3 Reported Observations

The extracted data with a summary of results from all 43 studies can be found in Table 3. 24 studies (n = 999 participants; percentage of women ranging from 16.7% [47] to 100% [48] with a mean age ranging from 30.3 to 55.1 years old) focused on acoustic (including paralinguistic) features [44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62]. 19 studies (n = 535 participants; percentage of women ranging from 0% to 75%, with a mean age between 21.7 and 68.0 years old) focused on lexical‐semantic features (e.g., semantic coherence, word type frequencies, and sentiment analysis) [42, 43, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75]. Manual coding procedures were more common among lexico‐semantic studies.

3.4 Distinguishing BD From Other Disorders

Lexico‐semantic features differentiated speech in BD from healthy individuals and patients with schizophrenia, but comparisons between bipolar mood states per se were limited. Most studies compared euthymic BD with other groups (healthy controls, SCZ, MDD); this approach assessed between‐group differences but not within‐BD differences. Various combinations of less coherent semantic clustering, wider vector angles, increased verbosity, and semantic fluency performance distinguished bipolar disorder from the other groups. Grouping of variables following Low et al. (2020) [41] showed that for the acoustic domain, the prosodic category (pauses, speech rate, pitch) was investigated the most, in 23 out of 24 studies, followed by the spectral domain (MFCCs) for 11 out of 24 studies. For the semantic domain, lexical features (specific word usage such as pronouns or verbs) were the most common group at 12 out of 19 studies, followed by syntactic (language structure) and coherence (flow of content across words and sentences), both used in 8 papers. For investigating noncategorical but individual key features in BD, see Table 2. Our results show more overlap in measures used in the acoustic compared to the semantic domain, both for categories and key variables.

3.5 Detecting and Predicting States Within BD

Acoustic features were most effective in differentiating between: (1) manic and euthymic states (for instance Faurholt‐Jepsen et al. [52]; Pan et al. [59]); (2) manic/hypomanic and depressed states (for instance Guidi et al. [48]; Maxhuni et al. [57]); and (3) manic, mixed, and depressed states [44, 53]; for full results see Table 3. While speech‐based features, such as jitter and pitch, were able to distinguish between the presence or absence of manic/depressed states (or in the case of Osmani [58], detecting state change), none of the studies reported on features that tracked changes in symptom severity or predicted a forthcoming relapse or switch.

3.6 Classification

Results of classifying BD from other groups ranged from an area under the curve (AUC) of 0.58 (distinguishing BD from MDD [46]) to an accuracy of 73% (distinguishing BD from healthy controls [64]). Classification or detection of BD states ranged from an AUC of 0.63 [45] to an AUC of 0.89 [52].

4 Discussion

To our knowledge, this is the first systematic review to examine markers of speech in individuals with BD. The studies we identified confirm the emerging interest in the use of speech features to study mood states in BD and report several observable linguistic patterns, specifically in the acoustic and lexical‐semantic domains across bipolar mood states.

We note that (1) distinct speech patterns can be recovered from individuals with BD compared to healthy controls and people with schizophrenia or MDD in some studies. Across many examined variables, the degree of impairment in the speech of individuals with BD occupies an intermediate position: being less severe than schizophrenia but also not as typical as healthy controls. This finding suggests a potential “ladder of severity” in speech abnormalities, positioning BD between healthy controls and schizophrenia [43, 66, 72, 73, 74, 76]. We note a tendency to use schizophrenia as the most common psychiatric comparator, with no studies contrasting BD and other diagnoses that are potential comorbidities (e.g., ADHD, PTSD, borderline personality, or substance‐induced mood states), highlighting a key research gap.

We also note that (2) the capacity of speech and language features to identify different states of BD. Acoustic parameters appear to have reasonable success in discriminating between manic/mixed states and depressive states in BD. Among non‐acoustic features, derailment, increased grammatical errors, word repetition, and verb use were associated with the manic state, while increased use of personal pronouns, modifying adverbs, and poorer verbal fluency related to the depressive state. These differential patterns also contributed to statistical separation between unipolar and bipolar depressed youth (who had more pressured speech) and between hypomanic and depressed mood states within BD.

However, for (3) predicting impending mood states, the current state of evidence linking speech patterns with BD mood states is marred by a notable lack of longitudinal, within‐subject examinations. Only 14 (33%) of studies in our review attempted repeated measurements in relation to mood states (see for instance Faurholt‐Jepsen et al. [52]; Pan et al. [59]), More recently, Kaczmarek‐Majer et al. [77] demonstrated that acoustic features can predict BD phases, achieving approximately 71% accuracy in classifying euthymia, depression and manic/mixed states/However, prediction of upcoming mood state changes, rather than concurrent classification, remains undemonstrated. Next to individual prediction, longitudinal, within‐individual design also reduces inter‐individual confounding effects [51] and allows evaluation of the individual predictive value of speech markers. These and other notable research gaps were appraised using Miles' Gap Framework (2017) [78] (see Table 4).

A promising development in recent times is the emergence of studies, for example, the PRIORI project [79] and CALIBER [80], as well as the Kaczmarek‐Majer et al. [77] study, that are leading these efforts. Notably, Kaczmarek‐Majer et al. [77] built on the methodological framework of Faurholt‐Jepsen et al. [52], adopting their cutoff thresholds and time‐windowing approaches (Kaczmarek‐Majer et al. [77], Faurholt‐Jepsen et al. [52]). However, such methodological continuity remaines the exception rather than the norm. Several acquisitions and analysis approaches are being used without external validation for data‐driven classification studies, with variables of interest not being selected on the basis of common theoretical or psychometric considerations. Sharing datasets and methodologies across research groups will promote collaboration and accelerate progress in the field. The representativeness of the BD samples studied to date is also poor; sociodemographic characteristics (age, parental socioeconomic status, education, native language status, and sex [8]), substance use [81], and treatment exposure [82] may influence speech features, but these are not explicitly accounted for in most studies. Similarly, integrating multiple analytic levels—semantic, acoustic, and temporal—will provide a richer understanding of speech changes linked to mood states.

Until these issues are addressed, speech markers are not ready as a clinical tool for relapse prediction or diagnostic decisions in BD. However, the state‐related acoustic and lexico‐semantic patterns identified in this literature suggest a near‐term role as an aid to clinical assessment, particularly when repeated speech samples can be compared within individuals over time (Hansen et al. [83]; Gumus et al. [84]). Such approach may help clinicians detect subtle emerging hypomanic or mixed features that are frequently missed, and prompt earlier, more targeted assessment (Anmella et al. [80]). Clinically, this could be the most valuable in the intervals between scheduled psychiatric appointments during which mood state changes can go undetected until they become severe (Cerimele and Fortney, [85]). Passive, smartphone‐based speech collection, as employed in Kaczmarek‐Majer et al. [77] and the PRIORI project, offer a feasible and relatively unobtrusive means of continuing monitoring in naturalistic settings (Kaczmarek‐Majer et al. [77], Ann Arbor M., [79]). Any clinical deployment will require prospective validation to false alarms, and safeguards for privacy and consent.

Our work has several strengths, including the pre‐registration of our search, methodical bias assessment of individual studies, and topic analysis to synthesize emerging trends. Various limitations must also be considered when interpreting our observations. We limited our search to published English language reports and adult samples (age > 18). We were not able to undertake a quantitative synthesis (meta‐analysis) due to the notable methodological heterogeneity among studies. Nonetheless, based on the lack of consensus regarding variable selection and acquisition protocols, we do not expect more homogenous observations to emerge from unpublished studies or studies published in other languages. A program of research that addresses the issues raised in Table 4 may lend itself to a quantitative synthesis in the future.

We note a potential for speech markers to aid in three key areas of need in bipolar disorder: (1) distinguishing BD from other disorders; (2) identifying different BD states, such as euthymia and mania; and (3) predicting impending mood states. While there is ample evidence supporting the pursuit of this line of work, the predictive value for each of the above clinical applications remains untested. Future research, particularly longitudinal studies, and concerted community efforts are needed to position speech assessments as a key approach in the study of BD.

Funding

This work was supported by Brain and Behavior Research Foundation, Canada First Research Excellence Fund, Canadian Institutes of Health Research and Fonds de Recherche du Québec—Santé.