The winning system in the Interspeech 2025 challenge on speech emotion recognition in naturalistic conditions reached a macro-F1 of just 0.4316 across eight emotions on real podcast speech, while the best of 10 benchmarked models scores 75.32% unweighted accuracy on the acted RAVDESS corpus (EmoBox benchmark, Interspeech 2024). The same split runs through clinical voice research: deep learning models pool to 0.83 sensitivity for detecting depression from speech (Lu et al., BMC Psychiatry 2025), yet only 2 of 68 mental health voice studies tested their model on an independent dataset (Uy et al., Frontiers in Digital Health 2026). Meanwhile, teletherapy is now routine: 12.3% of US adults, or 32.6 million people, received mental health treatment via telehealth in 2025 (SAMHSA, 2025 NSDUH). We aggregated data from the MSP Lab challenge leaderboards, the EmoBox and ParaLBench benchmarks, BMC Psychiatry, Annals of Family Medicine, JMIR, SAMHSA, FAIR Health, CDC NCHS, the American Psychological Association and the text of the EU AI Act. For the dementia side of voice health research, see our separate roundup on voice biomarkers for Alzheimer’s and cognitive impairment.
This page reports research findings only. It is not medical advice, and none of the tools described here replace an evaluation by a qualified clinician.
TL;DR
- Top macro-F1 on eight-class naturalistic speech emotion recognition was 0.4316, against a 0.3293 baseline (MSP Lab, Interspeech 2025 SER Challenge leaderboard).
- The best model reached 75.32% unweighted accuracy on acted RAVDESS but 22.24% on MSP-Podcast (EmoBox benchmark, Interspeech 2024).
- 17.98% of 267,905 MSP-Podcast speaking turns had no rater agreement on the primary emotion (Busso et al., MSP-Podcast Corpus 2025).
- Human listeners identified acted emotions in RAVDESS speech audio with 0.63 proportion correct (Livingstone and Russo, PLOS ONE 2018).
- Deep learning depression detection from speech pooled 0.83 sensitivity and 0.86 specificity across 16 studies (Lu et al., BMC Psychiatry 2025).
- A voice tool studied in 14,898 adults (4,456 in validation) reached 71.3% sensitivity and 73.5% specificity for moderate to severe depression (Mazur et al., Annals of Family Medicine 2025).
- Only 2 of 68 voice mental health studies used independent or cross-dataset validation (Uy et al., Frontiers in Digital Health 2026).
- 32.8% of US adults with any mental illness received treatment via telehealth in 2025 (SAMHSA, 2025 NSDUH).
- 52.1% of commercially insured telehealth patients had a mental health diagnosis in Q1 2026 (FAIR Health, Quarterly Telehealth Regional Tracker 2026).
- The EU AI Act ban on emotion inference at work and in schools has applied since 2 February 2025 (Regulation (EU) 2024/1689).
- 29% of US psychologists used AI at least monthly in 2025, up from 11% in 2024 (APA, 2025 Practitioner Pulse Survey).
1. Benchmark Accuracy on Standard Emotion Corpora
Three metrics recur in this section. Unweighted accuracy (UA) averages the hit rate of each emotion class, so every emotion counts equally. Weighted accuracy (WA) is the plain share of correct answers, which rewards a model for getting the most common emotion right. Macro-F1 is a similar class-balanced score that also penalizes false alarms.
On acted or tightly controlled corpora, the best model scores between 73.54% UA (IEMOCAP) and 84.62% UA (ESD). The best of 10 pre-trained models scored 84.62% UA on the ESD corpus but only 22.24% UA on MSP-Podcast, a corpus of real podcast speech (EmoBox benchmark, Interspeech 2024). On MSP-Podcast the weighted accuracy of the same model is roughly double its unweighted accuracy, a sign that the model does well on frequent emotions and poorly on rare ones. Headline accuracy figures from acted corpora therefore say little about how a system behaves in a phone call or a therapy session. The full per-dataset tables are in the EmoBox paper.
| Metric | Value | Source |
|---|---|---|
| Best UA among 10 models (Whisper large v3): IEMOCAP (acted dyadic dialogue) and RAVDESS (acted speech) | 73.54% and 75.32% | EmoBox (Ma et al.), Interspeech 2024 |
| RAVDESS corpus: professional actors and speech audio files | 24 actors (12 female, 12 male), 1,440 speech audio files | Livingstone and Russo, RAVDESS Zenodo record 2018 |
| CREMA-D and ESD (acted corpora), best UA | 76.75% and 84.62% | EmoBox (Ma et al.), Interspeech 2024 |
| MSP-Podcast (natural podcast speech), best model: UA vs WA | 22.24% vs 44.10% | EmoBox (Ma et al.), Interspeech 2024 |
| MELD (TV-series dialogue), best model: UA vs WA | 31.54% vs 51.89% | EmoBox (Ma et al.), Interspeech 2024 |
| IEMOCAP four-class emotion, best UA among 14 acoustic foundation models (Whisper-large) | 0.690 | ParaLBench benchmark, arXiv 2411.09349 (2024) |
| IEMOCAP four-class emotion, UA of WavLM-large | 0.685 | ParaLBench benchmark, arXiv 2411.09349 (2024) |
| Size of the IEMOCAP corpus | About 12 hours of audiovisual data | USC SAIL Lab, IEMOCAP database page |
Context: EmoBox covers 32 emotion datasets in 14 languages. The corpora differ in their number of emotion classes and in how the speech was produced (IEMOCAP’s categorical labels center on four emotions, RAVDESS speech uses eight), so scores are comparable within a row but not across rows. The ParaLBench results reach similar conclusions with a different model pool.
2. Naturalistic Conditions: The Interspeech 2025 Challenge
The Interspeech 2025 challenge asked teams to classify eight emotions (anger, happiness, sadness, fear, surprise, contempt, disgust, neutral) in spontaneous speech from the MSP-Podcast corpus, and to predict arousal, valence and dominance on a 1 to 7 scale. The winning macro-F1 of 0.4316 is a gain of about 0.10 over the 0.3293 baseline (our calculation from the official leaderboard), which is real progress but still means most eight-way decisions on natural speech are hard calls.
The dimensional task tells a similar story. CCC, the concordance correlation coefficient, measures how closely predicted scores track human ratings (1 means perfect agreement). Top systems track valence (how positive or negative a voice sounds) far better than dominance (how in-control a speaker sounds). The organizers’ summary paper notes that combining audio and text foundation models produced the most robust systems.
| Metric | Value | Source |
|---|---|---|
| Task 1 (eight emotion classes) baseline macro-F1 | 0.3293 | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Task 1 top macro-F1 (team NTUA) | 0.4316 | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Task 2 (arousal, valence, dominance) average CCC: baseline vs top system (team SAIL) | 0.5797 vs 0.6076 | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Task 2 top system: valence CCC vs dominance CCC | 0.6829 vs 0.498 | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Task 1 submissions reported by one participating team | 166 | Dutta et al., ABHINAYA, arXiv 2505.18217 (2025) |
| Same team’s system, scored by organizers after the deadline: macro-F1 | 44.02% | Dutta et al., ABHINAYA, arXiv 2505.18217 (2025) |
| Challenge training data: neutral samples vs fear samples | Neutral about 26 times more frequent | Dutta et al., ABHINAYA, arXiv 2505.18217 (2025) |
Context: the first-place MEDUSA system used an ensemble of 14 models yet was trained on a single RTX 3090 graphics card, so the ceiling here is the data, not compute. The challenge paired imbalanced training data with balanced test sets across emotion classes (Dutta et al., ABHINAYA 2025), which is why imbalance handling featured in the leading entries.
3. Label Ambiguity, Human Agreement and Cross-Corpus Transfer
The low scores in section 2 are partly a measurement problem. Nearly one in five speaking turns in the final MSP-Podcast release (48,180 of 267,905) had no plurality agreement on the primary emotion among at least five raters (Busso et al., The MSP-Podcast Corpus 2025). When people cannot agree on a label, a model scored against that label inherits the disagreement as a ceiling.
Transfer between corpora is the second weak point. A model trained on one dataset and tested on another (cross-corpus testing) faces new speakers, microphones and labeling habits. In EmoBox’s cross-corpus test, each corpus contributes 240 test clips balanced across four emotions (angry, happy, neutral, sad), so random guessing would score about 25%. The strongest model, Whisper large v3, ranged from 36.44% to 66.91% accuracy across the 12 train-test pairs.
| Metric | Value | Source |
|---|---|---|
| MSP-Podcast 2.0: annotated audio, speaking turns, speakers, raters per turn | 409 hours, 267,905 turns, 3,641 speakers, at least 5 raters per turn | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Speaking turns with no plurality agreement on the primary emotion | 48,180 of 267,905 (17.98%, our calculation) | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Inter-rater agreement on the primary emotion (kappa) | 0.411 | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Speaking turns labeled fear | 1,943 of 267,905 (0.73%, our calculation) | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Untrained human listeners naming the acted emotion in RAVDESS speech, audio only (proportion correct) | 0.63 (95% CI .62 to .64) | Livingstone and Russo, PLOS ONE 2018 (most recent available data) |
| Whisper large v3 cross-corpus accuracy: trained on IEMOCAP, tested on RAVDESS vs trained on RAVDESS, tested on SAVEE | 38.24% vs 66.91% | EmoBox cross-corpus leaderboard, 2024 |
| Valence agreement (CCC) of an annotator-mapped model: tested on MSP-Podcast vs transferred to IEMOCAP | .647 vs .516 | Tavernor and Mower Provost, arXiv 2509.12295 (2025) |
Context: the RAVDESS human figure and the model scores in section 1 are not a head-to-head contest, since listeners chose among eight labels plus a ‘none of these’ option under different conditions. The EmoBox cross-corpus numbers are on the public leaderboard, and the RAVDESS validation data is documented on its Zenodo record.
4. Voice-Based Depression and Anxiety Detection Accuracy
Clinical voice models usually do not classify discrete emotions. They predict whether a person’s score on a questionnaire such as the PHQ-9 (depression) or GAD-7 (anxiety) crosses a threshold. Pooled across 16 deep learning studies, speech-based depression detection reached 0.83 sensitivity and 0.86 specificity (Lu et al., BMC Psychiatry 2025). The largest single study below reports lower numbers, a reminder that pooled averages from many small samples can flatter.
That study collected at least 25 seconds of free speech from 14,898 adults in the US and Canada and, on its 4,456-sample validation set, reported 71.3% sensitivity (95% CI 69.0 to 73.5) and 73.5% specificity (95% CI 71.5 to 75.5) (Mazur et al., Annals of Family Medicine 2025). The study was funded through NSF grants awarded to the developer, and three authors are employees of that company. Anxiety models are far less studied, and a pregnancy study shows how fragile they are when only the patient’s own speech is kept.
| Metric | Value | Source |
|---|---|---|
| Deep learning speech models for depression: pooled sensitivity / specificity (16 studies) | 0.83 (95% CI 0.77 to 0.88) / 0.86 (95% CI 0.80 to 0.90) | Lu et al., BMC Psychiatry 2025 |
| Pooled AUC: deep learning vs traditional machine learning | 0.91 vs 0.89 | Lu et al., BMC Psychiatry 2025 |
| Voice tool vs PHQ-9 score of 10 or more (study of 14,898 adults, 4,456 in validation): sensitivity / specificity | 71.3% / 73.5% | Mazur et al., Annals of Family Medicine 2025 |
| Validation samples returned as ‘further evaluation recommended’ instead of a yes/no result | 920 of 4,456 (about 20%) | Mazur et al., Annals of Family Medicine 2025 |
| French adults from the general population (865 people): AUC for depression (PHQ-9) and anxiety (GAD-7) | 0.76 and 0.77 | Riad et al., Journal of Medical Internet Research 2024 |
| Same study: F1 score for PHQ-9 and GAD-7 risk | 0.49 and 0.50 | Riad et al., Journal of Medical Internet Research 2024 |
| Pregnancy study (146 women): GAD model F1, full interview vs only verified patient speech | 76% to 80% vs 51% to 65% | Bayrampour et al., Journal of Clinical Psychiatry 2026 |
| Voice pitch (fundamental frequency) in depression vs controls, pooled over 6 studies | 1.82 Hz lower, p = 0.56 (not significant) | Cantor-Cutiva et al., Depression and Anxiety 2026 |
Context: the AUC gap between deep learning and traditional models was not statistically significant in the BMC Psychiatry analysis, and the JMIR anxiety and depression study shows that a respectable AUC can still pair with a modest F1 when few people in the sample cross the threshold. For consumer apps that record voice diaries rather than screen for disease, see our roundup on AI voice journaling and therapy apps.
5. Evidence Gaps: Validation, Fairness and Clinical Readiness
The accuracy numbers in section 4 come mostly from models tested on data that looks like their training data. A 2026 scoping review found that just 2 of 68 voice and speech mental health studies (3%) validated their model on an independent or cross-dataset sample (Uy et al., Frontiers in Digital Health 2026). Read alongside section 3, where accuracy falls sharply across corpora, this is the single biggest reason to discount headline accuracy.
Fairness is the other open question. In the 14,898-person depression study, sensitivity for men was 14.7 percentage points lower than for women (59.3% vs 74.0%), and men were over-represented among missed cases (31.7% of false negatives vs 27.9% of the validation set). Speech technology has a long record of uneven performance across speaker groups, as our roundup on accent bias in speech recognition documents.
| Metric | Value | Source |
|---|---|---|
| Voice mental health studies with independent or cross-dataset validation | 2 of 68 (3%) | Uy et al., Frontiers in Digital Health 2026 |
| Studies in the same review that targeted depression | 56 of 68 (82%) | Uy et al., Frontiers in Digital Health 2026 |
| Studies that captured speech passively (ambient recording) | 3 of 68 (4%) | Uy et al., Frontiers in Digital Health 2026 |
| Voice screening dataset evaluations with external validation / real-world testing | 32% / 8% (of 37 evaluations) | Muntean-Codrea et al., Life 2026 |
| Dataset evaluations rated high clinical readiness | 11% | Muntean-Codrea et al., Life 2026 |
| Included studies (of 34) that addressed anxiety | 1 | Muntean-Codrea et al., Life 2026 |
| Clinical voice AI explainability studies with human-in-the-loop evaluation by stakeholders | 0 of 30 | Ebraheem et al., Journal of Medical Internet Research 2026 |
| Depression voice tool sensitivity: men vs women | 59.3% vs 74.0% | Mazur et al., Annals of Family Medicine 2025 |
Context: the Life review states that ‘no evaluation met all five indicators’ of clinical readiness (Muntean-Codrea et al., Life 2026). The JMIR explainability review found that 6 of its 30 studies used external data to check their explanations.
6. Teletherapy Use for Mental Health
Telehealth is where voice analysis would most plausibly run, because the session audio already passes through software. In 2025, 44.7% of US adults with a past-year major depressive episode, or 8.8 million people, received mental health treatment via telehealth (SAMHSA, 2025 NSDUH Annual National Report). Population-wide use edged down from 12.8% of adults in 2024 to 12.3% in 2025, which suggests teletherapy has settled into a stable channel rather than a pandemic spike.
Claims data show the same picture from the payer side: mental health is the top telehealth diagnostic category in every age group, and psychotherapy is the second most common telehealth procedure category (FAIR Health, Quarterly Telehealth Regional Tracker January to March 2026). Among psychologists, practice went from 64% fully remote in 2020 to 21% in 2023, with 67% working hybrid (APA, Practitioner Survey infographic 2023, most recent available data for this breakdown).
| Metric | Value | Source |
|---|---|---|
| US adults receiving mental health treatment via telehealth, 2025 | 12.3% (32.6 million) | SAMHSA, 2025 NSDUH Annual National Report |
| US adults receiving mental health treatment via telehealth, 2024 | 12.8% (33.4 million) | SAMHSA, 2024 NSDUH Annual National Report |
| Adults with any mental illness treated via telehealth, 2025 | 32.8% (17.9 million of 54.6 million) | SAMHSA, 2025 NSDUH Annual National Report |
| Telehealth patients (commercially insured) with a mental health diagnosis, Q1 2026 | 52.1% | FAIR Health, Quarterly Telehealth Regional Tracker 2026 |
| Same measure, patients aged 19 to 30 vs aged 65 and over | 69.4% vs 22.0% | FAIR Health, Quarterly Telehealth Regional Tracker 2026 |
| Telehealth patients billed for psychotherapy services | 35.0% | FAIR Health, Quarterly Telehealth Regional Tracker 2026 |
| Commercially insured patients with any telehealth claim: urban vs rural | 18.6% vs 10.3% | FAIR Health, Quarterly Telehealth Regional Tracker 2026 |
| Video-based vs face-to-face psychotherapy, 12 randomized trials (900 people): Hedges g for symptom reduction | -0.09 (95% CI -0.52 to 0.33), no significant difference | JMIR systematic review and multilevel meta-analysis 2026 |
Context: the NHIS, a different federal survey, found that 14.0% of US adults received counseling or therapy from a mental health professional in 2024, rising to 20.2% among adults aged 18 to 29 (CDC NCHS Data Brief No. 564). The video therapy result comes from the 2026 JMIR meta-analysis. For crisis lines rather than scheduled sessions, see our 988 Suicide and Crisis Lifeline call volume roundup.
7. Regulation, Ethics and Clinician Attitudes
Emotion recognition is one of the few AI uses the EU has banned outright in specific settings. Article 5(1)(f) of the EU AI Act prohibits AI systems that infer emotions of a natural person in the workplace and in education institutions, with an exception for medical or safety reasons, and this ban has applied since 2 February 2025 (Regulation (EU) 2024/1689). Outside those settings, emotion recognition systems are listed as high-risk, and anyone deploying one must inform the people exposed to it. The Act’s recitals define emotions as states such as happiness, sadness, anger, surprise and disgust, and exclude physical states such as pain or fatigue.
In the US, regulation targets AI therapy rather than emotion detection. Illinois signed the Wellness and Oversight for Psychological Resources Act on August 4, 2025, barring AI from therapeutic decision-making while allowing administrative support (Illinois IDFPR press release 2025). Clinicians are adopting AI quickly but with clear worries about privacy and bias (APA, 2025 Practitioner Pulse Survey).
| Metric | Value | Source |
|---|---|---|
| Date from which the EU ban on emotion inference at work and in education applies | 2 February 2025 | Regulation (EU) 2024/1689, Articles 5 and 113 |
| Maximum fine for prohibited AI practices | EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher | Regulation (EU) 2024/1689, Article 99 |
| General application date of the AI Act (emotion recognition listed as high-risk in Annex III) | 2 August 2026 | Regulation (EU) 2024/1689, Article 113 and Annex III |
| Illinois fine per confirmed violation of the AI therapy law | Up to $10,000 | Illinois IDFPR press release, August 2025 |
| US psychologists who had never used AI in their practice: 2024 vs 2025 | 71% vs 44% | APA, 2025 Practitioner Pulse Survey |
| Psychologists using AI at least monthly: 2024 vs 2025 | 11% vs 29% | APA, 2025 Practitioner Pulse Survey |
| Psychologists naming potential data breaches / biased input and output as an AI concern, 2025 | 67% / 63% | APA, 2025 Practitioner Pulse Survey |
| Psychologists with no openings for new patients / with a waitlist, 2025 | 46% / 40% | APA, 2025 Practitioner Pulse Survey |
Context: the Illinois Act passed unanimously in both chambers. The APA survey had 1,742 respondents from 26,502 delivered invitations, a 6.6% completion rate, so small shifts between years should be read with care.
Summary: Speech Emotion Recognition and Teletherapy by the Numbers
| Metric | Value | Source |
|---|---|---|
| Top eight-class macro-F1 on naturalistic speech (vs baseline) | 0.4316 (vs 0.3293) | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Top average CCC for arousal, valence and dominance | 0.6076 | MSP Lab, Interspeech 2025 SER Challenge leaderboard |
| Best UA on RAVDESS (acted) vs MSP-Podcast (natural) | 75.32% vs 22.24% | EmoBox (Ma et al.), Interspeech 2024 |
| MSP-Podcast turns with no rater agreement on the primary emotion | 17.98% of 267,905 | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Rater agreement on the primary emotion (kappa) | 0.411 | Busso et al., The MSP-Podcast Corpus, arXiv 2025 |
| Neutral vs fear samples in challenge training data | Neutral about 26 times more frequent | Dutta et al., ABHINAYA, arXiv 2025 |
| Human listeners on RAVDESS speech audio (proportion correct) | 0.63 | Livingstone and Russo, PLOS ONE 2018 |
| Deep learning depression detection: pooled sensitivity / specificity | 0.83 / 0.86 | Lu et al., BMC Psychiatry 2025 |
| Pooled AUC, deep learning depression models | 0.91 | Lu et al., BMC Psychiatry 2025 |
| Voice tool (14,898-adult study): sensitivity / specificity | 71.3% / 73.5% | Mazur et al., Annals of Family Medicine 2025 |
| Voice tool sensitivity: men vs women | 59.3% vs 74.0% | Mazur et al., Annals of Family Medicine 2025 |
| AUC for anxiety (GAD-7) from speech, 865 adults | 0.77 | Riad et al., Journal of Medical Internet Research 2024 |
| Voice mental health studies with independent validation | 2 of 68 (3%) | Uy et al., Frontiers in Digital Health 2026 |
| Dataset evaluations rated high clinical readiness | 11% | Muntean-Codrea et al., Life 2026 |
| US adults treated for mental health via telehealth, 2025 | 12.3% (32.6 million) | SAMHSA, 2025 NSDUH Annual National Report |
| Adults with a past-year major depressive episode treated via telehealth, 2025 | 44.7% (8.8 million) | SAMHSA, 2025 NSDUH Annual National Report |
| Telehealth patients with a mental health diagnosis, Q1 2026 | 52.1% | FAIR Health, Quarterly Telehealth Regional Tracker 2026 |
| EU ban on emotion inference at work and in education applies from | 2 February 2025 | Regulation (EU) 2024/1689 |
| Psychologists using AI at least monthly, 2025 | 29% | APA, 2025 Practitioner Pulse Survey |
Methodology and Sources
- EmoBox: Multilingual Multi-corpus Speech Emotion Recognition Toolkit and Benchmark (Ma et al., Interspeech 2024) and its cross-corpus leaderboard: intra-corpus and cross-corpus accuracy.
- ParaLBench: A Large-Scale Benchmark for Computational Paralinguistics over Acoustic Foundation Models (arXiv 2024): IEMOCAP results for 14 foundation models.
- USC SAIL Lab, IEMOCAP database: corpus size.
- MSP Lab, Interspeech 2025 Challenge on Speech Emotion Recognition in Naturalistic Conditions: baselines and leaderboards; organizers’ summary in Naini et al., Proc. Interspeech 2025.
- Dutta et al., ABHINAYA (arXiv 2505.18217, 2025) and MEDUSA (arXiv 2506.09556, 2025): participant system papers.
- Busso et al., The MSP-Podcast Corpus (arXiv 2509.09791, 2025): corpus size, rater agreement, label distribution.
- Tavernor and Mower Provost, annotator modeling for cross-corpus SER (arXiv 2509.12295, 2025).
- Livingstone and Russo, RAVDESS, PLOS ONE 2018 and the RAVDESS Zenodo record.
- Lu et al., BMC Psychiatry 2025; Mazur et al., Annals of Family Medicine 2025; Riad et al., Journal of Medical Internet Research 2024; Bayrampour et al., Journal of Clinical Psychiatry 2026; Cantor-Cutiva et al., Depression and Anxiety 2026.
- Uy et al., Frontiers in Digital Health 2026; Muntean-Codrea et al., Life 2026; Ebraheem et al., Journal of Medical Internet Research 2026.
- SAMHSA, 2025 NSDUH Annual National Report and 2024 NSDUH Annual National Report.
- FAIR Health, Quarterly Telehealth Regional Tracker, January to March 2026, United States.
- CDC NCHS Data Brief No. 564, Mental Health Treatment Among Adults: United States, 2024.
- Video-based vs face-to-face psychotherapy meta-analysis, Journal of Medical Internet Research 2026.
- APA, 2025 Practitioner Pulse Survey and APA telehealth location infographic 2020 to 2023.
- Regulation (EU) 2024/1689 (EU AI Act), Official Journal text: Article 3(39), Article 5(1)(f), Article 50(3), Article 99(3), Article 113, Annex III point 1(c), recital 18.
- Illinois IDFPR press release on HB1806, August 4, 2025.
- Data watch: benchmark scores are not comparable across corpora because label sets, class balance and metrics differ (UA, WA, macro-F1 and CCC measure different things). The ABHINAYA abstract reports 4th place among 166 submissions; its 44.02% score was computed by the organizers after the deadline and is not an official ranking. The MSP-Podcast paper gives neutral as about 28% of turns in its text, while its table counts (79,117 of 267,905) work out to 29.53%, and it gives 409 hours in the text vs 407 hours in a table; percentages marked ‘our calculation’ were computed from the table counts. The depression voice tool study was funded through grants to its developer, with three company employees as authors. The Riad et al. data are from 2024 and the RAVDESS human ratings from 2018, the most recent available for those measures. NSDUH (any treatment, self-report) and NHIS (counseling or therapy from a professional) measure different things, so their percentages should not be compared directly. FAIR Health covers commercially insured patients only and replaced its monthly tracker with a quarterly one in 2026, so older monthly figures are not directly comparable. The APA surveys had low completion rates (6.6% in 2025). EU AI Act dates are quoted from the Official Journal text of Regulation (EU) 2024/1689; any later amendment to its timelines should be checked against the current consolidated version.
Last updated: October 25, 2026. We update this roundup quarterly, and the next refresh is expected when FAIR Health publishes its next Quarterly Telehealth Regional Tracker and the MSP Lab posts results of its next speech emotion recognition challenge.