Not one of 68 voice and speech mental health studies reviewed in 2026 measured burnout or emotional exhaustion directly (Uy et al., Frontiers in Digital Health 2026). That gap frames everything below. What the research does show is narrower: fundamental frequency (F0, the basic pitch of the voice) rises under acute stress with a pooled standardized mean difference of 0.55 that shrinks to a non-significant 0.17 after correcting for publication bias (Veiga et al., Stress and Health 2025). The strongest call center field result we found is an AUC of 0.783 for predicting depressive symptoms in 62 Japanese workers (Tani et al., Journal of Occupational Health 2025), and in the European Union, inferring employees’ emotions with AI has been prohibited since February 2, 2025. For the workforce side of the problem (attrition, abuse, quit intentions), see our call center agent burnout and turnover statistics. We aggregated data from JMIR Mental Health, JAMIA, Scientific Reports, the Journal of Voice, the Journal of Occupational Health, the European Commission, the Hungarian data protection authority (NAIH) and the US National Institute of Justice, and separated laboratory evidence from field deployment throughout.
TL;DR
- 0 of 68 voice and speech mental health studies measured burnout directly, and only 2 of 68 (3%) validated on an independent dataset (Uy et al., Frontiers in Digital Health 2026).
- 0 of 13 real-workplace AI mental health monitoring studies used voice or speech data (Jack et al., PLOS Digital Health 2026).
- Acute stress raises F0 with a pooled SMD of 0.55 (95% CI 0.30 to 0.80), falling to 0.17 and non-significant after trim-and-fill (Veiga et al., Stress and Health 2025).
- F0 increased significantly in 15 of 19 stress studies, the most consistent acoustic marker (Schewski et al., PLOS One 2025).
- Speech-based depression detection: pooled highest accuracy 0.81 vs lowest accuracy 0.66 across 105 studies (Maran et al., JMIR Mental Health 2025).
- A call-audio model predicted depressive symptoms in 62 call center operators with AUC 0.783 (Tani et al., Journal of Occupational Health 2025).
- Call center operators report a career prevalence of voice problems of 33% to 68% (Nair et al., Journal of Voice systematic review).
- 70.5% of 173 Turkish call center operators referred for assessment had a confirmed occupational disease, most often vocal cord disorders (Altundas Hatman and Torun, La Medicina del Lavoro 2022).
- AI Act fines for prohibited practices reach EUR 35,000,000 or 7% of worldwide annual turnover, whichever is higher (Regulation (EU) 2024/1689, Article 99).
- A bank’s call-analysis software could not recognise emotion in 91.96% of cases, and the bank was fined HUF 250,000,000 (NAIH decision NAIH-85-3/2022).
1. What Stress Does to the Voice: F0, Jitter, Shimmer and Speech Rate
Paralinguistic analysis looks at how something is said rather than what is said. The four classic features are F0 (pitch), jitter (tiny cycle-to-cycle wobble in pitch), shimmer (cycle-to-cycle wobble in loudness) and speech rate. The laboratory evidence supports one of them clearly: pitch goes up under acute stress. F0 rose significantly in 15 of 19 studies in a 2025 systematic review of 38 articles (Schewski et al., PLOS One 2025), while jitter and shimmer mostly showed no association.
The effect is real but smaller and less stable than the headline number suggests. A 2025 meta-analysis of 10 studies and 583 participants pooled an F0 increase of SMD 0.55, but after statistical correction for missing negative studies the estimate dropped to 0.17 with a confidence interval crossing zero (Veiga et al., Stress and Health 2025). The signal is also strongest in spontaneous speech, which is exactly the kind of speech a headset captures, and weakest in read or standardized speech. A within-person experiment found the voice responds to physiological stress, not to bad mood alone: features moved during a stress test but not during a social exclusion task (Kappen et al., Scientific Reports 2024).
| Metric | Value | Source |
|---|---|---|
| Pooled F0 increase under acute stress (10 studies, 583 participants) | SMD 0.55 (95% CI 0.30 to 0.80), I² = 69.2% | Veiga et al., Stress and Health 2025 |
| Same pooled F0 effect after trim-and-fill correction for publication bias | SMD 0.17 (95% CI -0.15 to 0.49), p = 0.29 | Veiga et al., Stress and Health 2025 |
| F0 effect in spontaneous speech vs in standardized (read) speech | SMD 0.79 vs SMD 0.14 | Veiga et al., Stress and Health 2025 |
| Stress studies reporting a significant F0 increase | 15 of 19 | Schewski et al., PLOS One 2025 |
| Prosodic features vs voice quality features showing a stress relationship | 52 of 69 vs 18 of 37 examinations | Schewski et al., PLOS One 2025 |
| Speech rate change during a stress task, 66 participants | Increased, d = 1.03 (95% CI 0.65 to 1.42) | Kappen et al., Scientific Reports 2024 |
| Jitter change during the same stress task | Decreased, d = -0.62 (95% CI -1.00 to -0.24) | Kappen et al., Scientific Reports 2024 |
| Shimmer change during the same stress task | No significant change, p = 0.201 | Kappen et al., Scientific Reports 2024 |
Context note: in the Veiga meta-analysis, 5 of the 10 studies were rated at serious overall risk of bias, and the authors state that evidence is insufficient to establish F0 parameters for stress. Schewski et al. also found no consistent acoustic pattern for fear or anxiety.
2. Detection Accuracy: Depression and Fatigue Models in Benchmarks
Burnout overlaps with depression and exhaustion, so the depression literature is the nearest proxy for what a burnout model could achieve. The headline numbers look strong, but they depend on which result each paper chooses to report. Across 105 studies, the pooled best-reported accuracy was 0.81 while the pooled worst-reported accuracy of the same body of work was 0.66 (Maran et al., JMIR Mental Health 2025). That 15-point gap is the spread between a paper’s best configuration and its weakest one, before any model meets a new population.
Older approaches still hold up: a JAMIA meta-analysis found hand-engineered acoustic features beat end-to-end deep learning (Liu et al., JAMIA 2024). Fatigue shows the same small-sample pattern: a sleep-deprivation study reached up to 94% on vowel recordings but only 15 men took part (Frontiers in Cell and Developmental Biology 2022). For how emotion classifiers perform outside the lab in a clinical setting, see our speech emotion recognition accuracy statistics.
| Metric | Value | Source |
|---|---|---|
| Pooled highest reported accuracy, speech-based depression detection (105 studies) | 0.81 (95% CI 0.79 to 0.83) | Maran et al., JMIR Mental Health 2025 |
| Pooled lowest reported accuracy in the same review | 0.66 (95% CI 0.63 to 0.69) | Maran et al., JMIR Mental Health 2025 |
| Pooled highest reported sensitivity vs specificity | 0.84 vs 0.83 | Maran et al., JMIR Mental Health 2025 |
| Included studies at high risk of bias in at least one domain | 47.6% (50 of 105) | Maran et al., JMIR Mental Health 2025 |
| Pooled accuracy of deep learning on speech for depression (8 studies) | 0.87 (95% CI 0.81 to 0.93) | Liu et al., JAMIA 2024 |
| Pooled accuracy of models using handcrafted acoustic features | 0.89 (95% CI 0.81 to 0.97) | Liu et al., JAMIA 2024 |
| Pooled F0 difference, depressed vs non-depressed speakers (6 studies) | -1.82 Hz, p = 0.56 (not significant) | Cantor-Cutiva et al., Depression and Anxiety 2026 |
| Fatigue classification after 36 hours awake, multi-vowel vs segmented speech (15 participants) | Up to 94% vs 76% | Frontiers in Cell and Developmental Biology 2022 |
Context note: 52.4% of the 105 depression studies were conference papers, the most-used public dataset was DAIC-WOZ (33.3% of studies), and study-specific (handcrafted) datasets were the most common overall (53.3%). The Cantor-Cutiva et al. review rated only 1 of 17 studies as strong quality. Most recent available data for fatigue: Frontiers in Cell and Developmental Biology, 2022.
3. Burnout Is the Missing Label: The Lab-to-Field Gap
This is the section that matters most for anyone buying a “burnout prediction” feature. A 2026 scoping review of 68 voice biomarker studies found zero that measured burnout or emotional exhaustion, and only 2 of 68 (3%) tested on an independent dataset (Uy et al., Frontiers in Digital Health 2026). The authors conclude the evidence does not yet support passive or ambient mental health assessment in routine practice.
The workplace literature tells the same story from the other direction. A review restricted to AI mental health monitoring in real workplaces kept 13 studies, and none of them used voice (Jack et al., PLOS Digital Health 2026). One of the few large field-style voice studies, with 4,155 people recording on their own phones, found that no stress-linked feature was shared across gender, language and task (Topalian et al., Digital Biomarkers 2026). A model trained on one group, in other words, has no guarantee of transferring to the next team. The same validation problem runs through voice biomarkers in other conditions too, as our voice biomarker statistics for cognitive impairment show.
| Metric | Value | Source |
|---|---|---|
| Voice and speech mental health studies measuring burnout or emotional exhaustion | 0 of 68 | Uy et al., Frontiers in Digital Health 2026 |
| Studies validated on an independent or cross-dataset sample | 2 of 68 (3%) | Uy et al., Frontiers in Digital Health 2026 |
| Studies using controlled elicitation (lab tasks) rather than natural speech | 54 of 68 (79%) | Uy et al., Frontiers in Digital Health 2026 |
| Stress-linked voice features shared between women and men, 679 adults with diabetes recording at home | None (25 features found in each sex) | Bour et al., PLOS Digital Health 2026 |
| Real-workplace AI mental health monitoring studies that used voice or speech | 0 of 13 | Jack et al., PLOS Digital Health 2026 |
| Real-workplace AI monitoring studies targeting burnout vs targeting stress | 1 of 13 vs 10 of 13 | Jack et al., PLOS Digital Health 2026 |
| Stress studies that were experimental (lab) vs field studies | 30 vs 7 articles | Schewski et al., PLOS One 2025 |
| Momentary-stress voice study on personal devices: features shared across gender, language and task | None (4,155 participants) | Topalian et al., Digital Biomarkers 2026 |
Context note: field data collection is possible. The ICASSP 2021 study by Gaballah et al. followed 144 hospital workers over a 10-week period with wearable audio and daily self-reported stress, but it targeted stress, not burnout (most recent available data: IEEE ICASSP, 2021). The diabetes cohort in the table above (Bour et al., PLOS Digital Health 2026) also found no feature shared between stress and diabetes-related distress, a reminder that “stress” in one study is not the same construct as in the next.
4. Call Center Deployments: What Field Studies Report
The closest thing to a real burnout-adjacent deployment comes from Japan, where one research group tested a commercial voice analysis system on operators’ actual customer calls. The model predicted depressive symptoms over 12 weeks with an AUC of 0.783 (95% CI 0.691 to 0.875), but only 62 of 240 recruited operators ended up in the analysis, the system is a proprietary black box, and the study was funded by its vendor (Tani et al., Journal of Occupational Health 2025). A companion randomized trial used the same tool in group sessions and found modest reductions in depression scores (Tani et al., Journal of Occupational Health 2024).
Two newer Korean datasets show the alternatives. A 400-worker diary study found that the words agents used tracked depressive symptoms better than their own mood ratings, which is language analysis rather than paralinguistics (Kim et al., Psychiatry Investigation 2026). A lab simulation with 31 call agents and actors playing angry customers reported results only for audio combined with heart, skin and brain sensors, so the voice’s own contribution is unknown (Lee et al., Scientific Data 2026).
| Metric | Value | Source |
|---|---|---|
| AUC for predicting depressive symptoms (CES-D 16 or higher) from call audio, 62 operators | 0.783 (95% CI 0.691 to 0.875) | Tani et al., Journal of Occupational Health 2025 |
| AUC for telling normal calls from high-workload (aggressive customer) calls in a simulated call center, audio combined with physiological sensors, 31 agents | 0.946 | Lee et al., Scientific Data 2026 |
| Operators with depressive symptoms at baseline in that cohort | 30.6% (19 of 62) | Tani et al., Journal of Occupational Health 2025 |
| Change in depression score at 12 weeks vs control: voice-analysis group sessions (5 or more) vs one-on-one supervisor meetings | -2.36 vs -1.47 CES-D points | Tani et al., Journal of Occupational Health 2024 |
| Korean emotional labour workers with depressive symptoms (PHQ-9 of 5 or higher), 400 analyzed | 50.8% (203 of 400) | Kim et al., Psychiatry Investigation 2026 |
| AUC of language-based sentiment model by day 10 of voice diaries | 75.1% | Kim et al., Psychiatry Investigation 2026 |
| AUC for perceived stress in a simulated call center, audio combined with physiological sensors, 31 agents | 0.783 | Lee et al., Scientific Data 2026 |
| Share of deceptive answers detected by Layered Voice Analysis in a US jail validation, 319 arrestees | 21.4% (specificity 95.2%) | US National Institute of Justice, Damphousse et al. 2007 |
Context note: the Japanese system used in both Tani studies is built on Layered Voice Analysis (LVA). An older independent test funded by the US National Institute of Justice found LVA and a rival voice stress analyzer were “not able to detect deception at a rate any better than chance” (NIJ report, 2007). That test measured lying, not mood, so it does not settle the depression question, but it is a reason to demand independent validation (most recent available data: NIJ, 2007).
5. Vocal Load: Voice Disorders in Call Center Work and Their Link to Burnout
Before voice is a sensor, it is a work tool, and the tool wears out. A Journal of Voice systematic review of 15 studies put the career prevalence of voice problems among call center operators at 33% to 68%, with a point prevalence of 27% (Nair et al., Journal of Voice). Hoarseness was the most common symptom, followed by vocal fatigue. That matters for any acoustic model: a hoarse agent’s jitter and shimmer shift for physical reasons, which a stress or burnout classifier could misread.
The clinical end of the spectrum is stark. Among 173 Turkish call center operators referred to an occupational disease hospital, 70.5% had a confirmed occupational disease, and of those later interviewed, 43.8% had been dismissed (Altundas Hatman and Torun, La Medicina del Lavoro 2022). Research on teachers, another voice-heavy profession, links burnout and voice disorders directly, so the two problems travel together. Our acoustic voice analysis and dysphonia statistics cover how clinics measure that damage.
| Metric | Value | Source |
|---|---|---|
| Career prevalence of voice problems in call center operators (15 studies) | 33% to 68% | Nair et al., Journal of Voice systematic review |
| Point prevalence of voice problems in call center operators | 27% | Nair et al., Journal of Voice systematic review |
| Turkish call center employees reporting voice problems since starting the job (264 surveyed) | 57.6% | Demir et al., PeerJ 2025 |
| Same employees reporting stress at work | 58.3% | Demir et al., PeerJ 2025 |
| Referred call center operators with a confirmed occupational disease (173 cases) | 70.5% | Altundas Hatman and Torun, La Medicina del Lavoro 2022 |
| Vocal cord nodules among female vs male operators with an occupational disease | 92.9% vs 62.4% | Altundas Hatman and Torun, La Medicina del Lavoro 2022 |
| Mean share of the working day spent phonating, 92 call center operators | 14.74% (range 4% to 31%) | Cantarella et al., Journal of Voice 2014 |
| Odds of a probable voice disorder in teachers positive on 2 to 4 burnout subscales vs 0 to 1 | OR 4.01, P = 0.013 | de Brito Mota et al., Journal of Voice 2019 |
Context note: in the Turkish PeerJ survey (Demir et al. 2025), stress at work was significantly associated with worse Voice Handicap Index scores (p < 0.001). Among 208 teachers in the de Brito Mota study, 64.4% were likely to have a voice disorder, and in a 2024 Galician study of 90 teachers, 16% showed symptoms compatible with burnout. Most recent available data on phonation time: Cantarella et al., Journal of Voice, 2014.
6. Legal and Ethical Limits: EU AI Act Article 5 and Regulator Action
Even if the science matured tomorrow, the legal room for employee emotion analysis in Europe has already closed. Article 5(1)(f) of the AI Act prohibits “the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions”, with an exception only for medical or safety reasons, and it has applied since February 2, 2025 (Regulation (EU) 2024/1689, Article 5). The European Commission’s guidelines, published February 4, 2025 and non-binding, close the obvious loophole: according to a Future of Privacy Forum analysis of the guidelines, the medical exception “does not extend to general well-being monitoring, such as stress or burnout detection” (FPF, March 24, 2026).
The ban targets inference from biometric data such as the voice. Analysing customers’ emotions is not covered, and inferring emotions from written text sits outside the prohibition, so a transcript-based model and a voice-based model can land on opposite sides of the line. Data protection law already bit before the AI Act did: Hungary’s authority fined a bank whose call-analysis software scored callers’ emotional state and also monitored its own employees’ voices (NAIH decision NAIH-85-3/2022).
| Metric | Value | Source |
|---|---|---|
| Date the workplace emotion-inference ban started applying in the EU | February 2, 2025 | Regulation (EU) 2024/1689 (AI Act), Article 5 |
| Maximum fine for a prohibited AI practice | EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher | Regulation (EU) 2024/1689 (AI Act), Article 99(3) |
| Publication of the Commission guidelines on prohibited practices (non-binding) | February 4, 2025, page last updated July 31, 2025 | European Commission, Guidelines on prohibited AI practices 2025 |
| Stress or burnout detection covered by the medical exception | No, general well-being monitoring is excluded | European Commission guidelines, as analysed by Future of Privacy Forum 2026 |
| Fine on a Hungarian bank for AI emotion analysis of call recordings | HUF 250,000,000 | NAIH decision NAIH-85-3/2022 |
| Share of cases where the bank’s software could not recognise emotion | 91.96% | NAIH decision NAIH-85-3/2022 |
| Retention of call audio files inside the voice analytics software | 45 days | NAIH decision NAIH-85-3/2022 |
| Deadline to prove compliance or stop the emotion analysis | 60 days after the decision became final | NAIH decision NAIH-85-3/2022 |
Context note: the European Commission guidelines page states that authoritative interpretation is reserved for the Court of Justice of the European Union. The NAIH decision also found that the software could recognise “the emotions of an employee and not that of the calling party”, the kind of workplace emotion inference that Article 5(1)(f) now prohibits.
Summary: Paralinguistic Voice Analysis and Agent Burnout Prediction by the Numbers
| Metric | Value | Source |
|---|---|---|
| Voice and speech mental health studies measuring burnout directly | 0 of 68 | Uy et al., Frontiers in Digital Health 2026 |
| Real-workplace AI mental health monitoring studies using voice | 0 of 13 | Jack et al., PLOS Digital Health 2026 |
| Pooled F0 increase under acute stress | SMD 0.55 (95% CI 0.30 to 0.80) | Veiga et al., Stress and Health 2025 |
| Pooled F0 increase after publication-bias correction | SMD 0.17, not significant | Veiga et al., Stress and Health 2025 |
| Stress studies with a significant F0 increase | 15 of 19 | Schewski et al., PLOS One 2025 |
| Speech rate effect under a lab stress task | d = 1.03 | Kappen et al., Scientific Reports 2024 |
| Depression detection, pooled highest vs lowest reported accuracy | 0.81 vs 0.66 | Maran et al., JMIR Mental Health 2025 |
| Deep learning depression detection, pooled accuracy | 0.87 | Liu et al., JAMIA 2024 |
| Handcrafted acoustic feature models, pooled accuracy | 0.89 | Liu et al., JAMIA 2024 |
| Call-audio model AUC for depressive symptoms, 62 operators | 0.783 | Tani et al., Journal of Occupational Health 2025 |
| Emotional labour workers with depressive symptoms (Korea) | 50.8% | Kim et al., Psychiatry Investigation 2026 |
| Career prevalence of voice problems in call center operators | 33% to 68% | Nair et al., Journal of Voice systematic review |
| Turkish call center employees with voice problems since starting the job | 57.6% | Demir et al., PeerJ 2025 |
| Referred operators with a confirmed occupational disease | 70.5% | Altundas Hatman and Torun, La Medicina del Lavoro 2022 |
| Maximum AI Act fine for prohibited practices | EUR 35,000,000 or 7% of turnover | Regulation (EU) 2024/1689, Article 99 |
| Hungarian bank fine for call emotion analysis | HUF 250,000,000 | NAIH decision NAIH-85-3/2022 |
Methodology and Sources
- Uy, Shaaban and Arnaout: Voice and speech as digital biomarkers of mental health, Frontiers in Digital Health 2026
- Jack et al.: Mental health monitoring in the workplace using artificial intelligence, PLOS Digital Health 2026
- Veiga et al.: The fundamental frequency of voice as a potential stress biomarker, Stress and Health 2025
- Schewski et al.: Measuring negative emotions and stress through acoustic correlates in speech, PLOS One 2025
- Kappen et al.: Acoustic and prosodic speech features reflect physiological stress, Scientific Reports 2024
- Topalian et al.: Voice signatures of momentary psychological stress in real-life environments, Digital Biomarkers 2026
- Bour et al.: Voice changes associated with stress and distress in people living with diabetes, PLOS Digital Health 2026
- Maran et al.: Performance of automatic speech analysis in detecting depression, JMIR Mental Health 2025
- Liu et al.: Diagnostic accuracy of deep learning using speech samples in depression, JAMIA 2024
- Cantor-Cutiva et al.: The role of voice acoustics in depression assessment, Depression and Anxiety 2026
- A rapid, non-invasive method for fatigue detection based on voice information, Frontiers in Cell and Developmental Biology 2022
- Gaballah et al.: Context-aware speech stress detection in hospital workers using Bi-LSTM classifiers, IEEE ICASSP 2021
- Tani et al.: Early detection model for depressive symptoms using voice emotion analysis among call center operators, Journal of Occupational Health 2025
- Tani et al.: Organizational intervention using Layered Voice Analysis, randomized controlled trial, Journal of Occupational Health 2024
- Kim et al.: Daily language as an objective indicator of depressive affective states in emotional labor workers, Psychiatry Investigation 2026
- Lee et al.: A multimodal dataset for assessing emotion, stress, and emotional workload in interpersonal work scenario, Scientific Data 2026
- Damphousse et al., US National Institute of Justice: Assessing the validity of voice stress analysis tools in a jail setting, 2007
- Nair et al.: Prevalence of voice problems, self-reported vocal symptoms and associated risk factors in call center operators, Journal of Voice
- Demir et al.: Self-reported voice problems in call center employees, PeerJ 2025
- Altundas Hatman and Torun: Occupational diseases among call center operators needing vocal rehabilitation, La Medicina del Lavoro 2022
- Cantarella et al.: Ambulatory phonation monitoring in a sample of 92 call center operators, Journal of Voice 2014
- de Brito Mota et al.: Voice disorder and burnout syndrome in teachers, Journal of Voice 2019; Garcia-Real et al.: Vocal problems and burnout syndrome in nonuniversity teachers in Galicia, Folia Phoniatrica et Logopaedica 2024
- Regulation (EU) 2024/1689 (AI Act): Article 5 and Article 99
- European Commission: Guidelines on prohibited artificial intelligence practices, 2025
- Future of Privacy Forum: Red lines under EU AI Act, emotion recognition in the workplace, 2026
- Hungarian National Authority for Data Protection and Freedom of Information (NAIH): Decision NAIH-85-3/2022 on AI-based speech signal processing
- Data watch: no peer-reviewed study we found reports an AUC for voice-based burnout detection against a validated burnout inventory, so every accuracy figure in this roundup refers to stress, fatigue or depression, not burnout itself. Both Tani et al. studies were funded by ES Japan, the vendor of the voice analysis system, with company employees as co-authors, and the 2025 prediction model was validated only within its own cohort. The Kim et al. study analysed transcribed words, not acoustic features. The NIJ figures measure deception, not mood, and date from 2007. Internal inconsistencies in the sources themselves: Altundas Hatman and Torun give 62.4% for vocal cord nodules in men in the text and 68.4% in Table 4; the fatigue study’s abstract (94% unitary voice, 81% long speech) does not fully match its results section (88% to 94% for vowels, 76% for segmented speech); Jack et al. state both 8 of 11 and 7 of 11 studies above 75% accuracy, so we did not use that figure. The Nair et al. review was published online in 2021 and printed in the Journal of Voice in 2024. The AI Act prohibition text is cited from the artificialintelligenceact.eu reproduction of the Official Journal text.
Last updated: October 27, 2026. We update this roundup quarterly, and the next refresh is expected when the next systematic review of voice biomarkers for workplace mental health is indexed in PubMed Central.