Emotion recognition
Emotion recognition, also called automatic emotion recognition or affect recognition, is the use of computational methods to identify or infer a person's emotional state from signals such as facial movements, the voice, written language, body movement and physiological measurements.[1][2] It grew out of affective computing, a research area that Rosalind Picard of the MIT Media Lab described in 1995 as computing "that relates to, arises from, or deliberately influences emotions."[3] Systems are typically machine learning models trained on labeled examples, using computer vision for faces, speech models for the voice, natural language processing for text, and signal processing for measurements such as brain activity.[4][5][6][7]
The field is contested. A 2019 review led by the psychologist Lisa Feldman Barrett concluded that the way people express anger, disgust, fear, happiness, sadness and surprise "varies substantially across cultures, situations, and even across people within a single situation," challenging the common assumption that emotional states can be readily inferred from facial movements.[8] Microsoft announced in 2022 that it would retire emotion inference from its Azure Face service,[9] the UK Information Commissioner's Office warned organisations against "immature" emotion analysis the same year,[10] and the EU AI Act has banned AI systems that infer emotions in workplaces and education institutions since 2 February 2025.[11] In September 2026 California enacted a ban, taking effect in January 2027, on employers using AI workplace surveillance tools to recognize or infer employees' emotional states.[12][42][43]
Definitions and scope
Definitions differ between law and research. The EU AI Act defines an "emotion recognition system" as an AI system for "identifying or inferring emotions or intentions of natural persons on the basis of their biometric data."[13] The Act's recital 18 lists emotions such as happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction and amusement, and excludes physical states such as pain or fatigue (for example, systems that detect fatigue in drivers or pilots) as well as the "mere detection of readily apparent expressions," such as noticing a smile or a raised voice without inferring an emotion from it.[14] Because the legal definition depends on biometric data, the European Commission's guidelines treat an AI system that infers emotions from written text as outside the workplace ban, while one that infers emotions from keystrokes, facial expressions, body postures or movements is inside it.[15]
Research usage is broader. Saif Mohammad's ethics sheet for automatic emotion recognition (AER) covers emotion recognition and sentiment analysis together, including text-based systems.[1] The AI Now Institute's 2019 report used "affect recognition" for technology that "claims to detect things such as personality, emotions, mental health, and other interior states based on physiological measurements such as facial expression, voice and gait."[2] Some shared tasks combine the two: SemEval-2018 Task 1, for example, combined emotion intensity, valence and emotion classification subtasks.[16]
History
Psychological foundations
A long-running line of research in psychology holds that some facial expressions signal discrete emotions across cultures. In a 1971 study, Paul Ekman and Wallace Friesen told members of a preliterate culture in New Guinea a short story, showed them three faces and asked them to pick the face that matched the emotion in the story; they reported that the results supported the hypothesis that associations between particular facial muscle patterns and discrete emotions are universal.[17] Ekman and Friesen published the Facial Action Coding System (FACS) in 1978, an anatomically based scheme that breaks facial expressions into "action units," the individual components of muscle movement; a revised edition appeared in 2002.[18] The target expression in each sequence of the CK+ facial dataset, for example, is fully FACS coded.[19]
A second tradition describes affect along continuous dimensions. James Russell's 1980 circumplex model of affect is organized by two dimensions, valence and arousal.[20][21] Datasets such as AffectNet and IEMOCAP label data both ways, with discrete categories and with valence and arousal (IEMOCAP also labels dominance).[22][23]
Affective computing
Picard's MIT Media Lab technical report "Affective Computing" (report no. 321, revised November 1995) framed the recognition of human affect as a pattern recognition problem and proposed applications in computer-assisted learning, perceptual information retrieval, the arts and entertainment, and human health.[3] Her book Affective Computing was published by MIT Press in 1997.[24] In a 2001 study, Picard, Elias Vyzas and Jennifer Healey recorded four physiological signals from one subject who tried to experience each of eight emotional states daily over several weeks; their best method reached 81 percent accuracy on the eight classes, and they reported that features of different emotions on the same day clustered more tightly than features of the same emotion on different days.[25] The IEEE Transactions on Affective Computing began publication in 2010; its first issue included an article by Picard titled "Affective Computing: From Laughter to IEEE."[26]
Benchmarks and deep learning
Early datasets were small and often posed.[22][27] The Cohn-Kanade database, released in 2000, became a widely used test bed, but its 2010 successor paper noted that its emotion labels "refer to what was requested rather than what was actually performed"; the Extended Cohn-Kanade (CK+) dataset revised and validated those labels and added non-posed smiles.[19] The INTERSPEECH 2009 Emotion Challenge was organized because speech emotion research lacked standardized corpora and test conditions; it used the FAU Aibo Emotion Corpus with fixed, speaker-independent training and test partitions, and its organizers wrote that the community needed "more spontaneous and less prototypical data."[27]
The Facial Expression Recognition 2013 (FER-2013) challenge, held on Kaggle for an ICML 2013 workshop organized by Ian Goodfellow, Dumitru Erhan and Yoshua Bengio, introduced a dataset of 35,887 grayscale face images of 48 by 48 pixels collected through Google image search.[4] The winning entry, by Yichuan Tang, was a convolutional neural network trained with an L2-SVM loss and reached 71.162 percent accuracy on the private test set; the organizers noted that convolutional networks beat hand-designed features such as SIFT, "but the difference in accuracy is not extreme."[4] Later datasets scaled up: AffectNet (2017) gathered more than one million facial images from the web using 1,250 emotion-related keywords in six languages, and about half were manually annotated.[22] In speech, self-supervised models such as WavLM are evaluated on emotion recognition as one of many downstream tasks in the SUPERB benchmark.[5][28]
Modalities and methods
Facial expressions
Facial emotion recognition systems classify a face into a fixed set of expression categories, predict continuous values such as valence and arousal, or both.[22] The construction of FER-2013 shows a typical pipeline: a face detector produced bounding boxes, human labelers rejected wrong images and corrected crops, and the images were resized and mapped onto seven categories (anger, disgust, fear, happiness, sadness, surprise and neutral).[4] The category counts in FER-2013 were uneven, from 547 disgust images to 8,989 happiness images.[4] Goodfellow estimated human accuracy on FER-2013 at 65 percent (plus or minus 5) and at 68 percent (plus or minus 5) on a separate set of 1,500 images of lab members acting out the seven expressions.[4] An alternative to predicting emotions directly is to detect FACS action units, which describe visible facial movements rather than emotions.[18][19]
Speech
Speech emotion recognition infers emotion from properties of the speech signal.[29] The Geneva Minimalistic Acoustic Parameter Set (GeMAPS), proposed in 2016 by Florian Eyben, Klaus Scherer, Björn Schuller and colleagues, selected a small standard set of voice parameters partly for their potential to index affective physiological changes in voice production, and was released with the openSMILE toolkit.[30] The SUPERB benchmark, introduced in 2021, tests whether representations from self-supervised speech models such as wav2vec 2.0 and HuBERT, whose pre-training had been explored mainly for speech recognition, also work for other speech tasks.[5][28] In SUPERB, the emotion recognition task uses IEMOCAP, drops the less frequent classes to keep four (neutral, happy, sad and angry), and reports accuracy with five-fold cross-validation.[5]
Text
Text-based emotion recognition is closely related to sentiment analysis. SemEval-2018 Task 1 ("Affect in Tweets") provided English, Arabic and Spanish tweets for five subtasks: emotion intensity regression, emotion intensity ordinal classification, valence regression, valence ordinal classification and emotion classification; 75 teams took part.[16] The GoEmotions dataset (2020) labels 58,000 English Reddit comments with 27 emotion categories or neutral; its BERT baseline reached an average F1 score of 0.46, which the authors said left "much room for improvement."[6]
Physiological signals
Physiological approaches use signals such as electroencephalography (EEG) and peripheral physiological measurements. The DEAP dataset (2012) recorded EEG and peripheral physiological signals from 32 participants as each watched 40 one-minute music video excerpts and rated them for arousal, valence, liking, dominance and familiarity; frontal face video was also recorded for 22 participants.[7] Picard's 2001 study had already found that such signals vary from day to day for the same person and emotion, which complicates recognition.[25]
Multimodal recognition
Several datasets combine face, voice and words. IEMOCAP contains about 12 hours of audiovisual data, including video, speech, facial motion capture and transcriptions, from actors performing improvised or scripted dyadic scenarios.[23] RAVDESS (2018) has 7,356 recordings from 24 professional actors, gender balanced, in face-and-voice, face-only and voice-only formats, each rated 10 times for emotional validity, intensity and genuineness.[31] MELD (2019) has about 13,000 utterances from 1,433 dialogues of the TV series Friends, labeled for emotion and sentiment in audio, visual and text modalities, for the task of emotion recognition in conversations.[32] Such corpora are used to train and evaluate multimodal models that combine the modalities.[32]
Datasets
| Dataset | Year | Modality | Size | Labels |
|---|---|---|---|---|
| CK+ | 2010 | Face (posed video sequences, some non-posed smiles) | Extends the 2000 Cohn-Kanade set | FACS action units, validated emotion labels[19] |
| IEMOCAP | 2008 | Audio, video, facial motion capture, text | About 12 hours | Categories plus valence, activation, dominance[23][44] |
| DEAP | 2012 | EEG, peripheral physiology, face video | 32 participants, 40 clips each | Arousal, valence, liking, dominance, familiarity[7] |
| FER-2013 | 2013 | Face images (48x48 grayscale) | 35,887 images | 7 categories[4] |
| AffectNet | 2017 | Face images from the web | More than 1,000,000 images, about half annotated | 7 categories plus valence and arousal[22] |
| RAVDESS | 2018 | Speech, song, face video | 7,356 recordings, 24 actors | Calm, happy, sad, angry, fearful, surprise, disgust (speech), neutral[31] |
| SemEval-2018 Task 1 | 2018 | Tweets (English, Arabic, Spanish) | Shared task, 75 teams | Emotion intensity, valence, emotion classes[16] |
| MELD | 2019 | Audio, video, text (TV dialogue) | About 13,000 utterances | Emotion and sentiment[32] |
| GoEmotions | 2020 | Reddit comments (English) | 58,000 comments | 27 emotions or neutral[6] |
Several of these datasets use acted or requested expressions (CK+, RAVDESS, IEMOCAP).[19][31][23] FER-2013's labels came from the emotion keywords used to find the images, checked by human labelers, and AffectNet's from manual annotation of web images.[4][22] The CK+ authors' remark that the original Cohn-Kanade labels described what was requested rather than what was performed is one example of the gap between a dataset label and a person's state.[19] Barrett and colleagues called for research on how people actually move their faces to express emotion in the contexts of everyday life.[8]
Applications
Proposed and actual uses include education, human-computer interaction, health, advertising and media testing, customer service, hiring and in-car monitoring. Picard's 1995 report sketched applications in learning, information retrieval, entertainment and health.[3] GoEmotions' authors cited empathetic chatbots and detection of harmful online behavior.[6] The Commission's AI Act guidelines discuss call centres that analyze callers' voices, cameras in shops and banks, and emotion-aware online advertising, including "intelligent billboards."[15]
In May 2021 Smart Eye, a Swedish maker of eye-tracking and automotive driver monitoring systems, agreed to acquire Affectiva, which spun out of the MIT Media Lab in 2009, for $73.5 million. The announcement said Affectiva's technology analyzed consumers' emotional reactions to brand content and that its interior sensing AI was used by car manufacturers to understand drivers and occupants.[33] In recruiting, the US company HireVue analyzed candidates' recorded video interviews, including their facial expressions, until it dropped facial analysis in January 2021 (see below).[34]
Criticism
Scientific validity
The central scientific objection is that facial movements are not reliable indicators of emotional states. Barrett and colleagues' 2019 review in Psychological Science in the Public Interest examined what it called the "common view" that emotional states can be readily inferred from facial movements, focusing on anger, disgust, fear, happiness, sadness and surprise.[8] It found that people do sometimes smile when happy, frown when sad and scowl when angry, more often than chance would predict, but that the same configuration of facial movements can express more than one emotion and often communicates something other than an emotional state.[8] The authors wrote that the common view shapes legal judgments, policy decisions, national security protocols and educational practices, as well as research in AI and computer vision.[8]
Other researchers report more regularity. A study by Alan Cowen, Dacher Keltner and colleagues, published in Nature in December 2020, used deep neural networks to examine 16 kinds of facial expression across thousands of contexts in 6 million videos from 144 countries, and found that each expression's associations with contexts were 70 percent preserved across 12 world regions.[35] The study examined associations between expressions and social contexts such as weddings or sporting competitions.[35]
Recital 44 of the EU AI Act states that "there are serious concerns about the scientific basis of AI systems aiming to identify or infer emotions," citing variation across cultures, situations and individuals, limited reliability, lack of specificity and limited generalisability.[36]
Bias, privacy and power
Critics also raise discrimination and privacy concerns. Recital 44 says such systems may lead to discriminatory outcomes and be intrusive, particularly where there is an imbalance of power, as in employment and education.[36] The ICO said in October 2022 that the algorithms were unable to detect emotional cues effectively, risking systemic bias, inaccuracy and discrimination; deputy commissioner Stephen Bonner said: "Developments in the biometrics and emotion AI market are immature. They may not work yet, or indeed ever."[10] SemEval-2018 Task 1 included an analysis of whether participating systems showed consistent bias toward a particular race or gender.[16] Mohammad's 2022 ethics sheet in Computational Linguistics sets out fifty ethical considerations for automatic emotion recognition, with special attention to privacy and to effects on social groups, and notes that such systems can help public health and commerce but can also be used to suppress dissidents and manipulate voters.[1] For broader context see algorithmic bias and AI ethics.
The AI Now Institute's 2019 report put the first of its twelve recommendations this way: "Regulators should ban the use of affect recognition in important decisions that impact people's lives and access to opportunities." It added that until then, AI companies should stop deploying it.[2]
Industry responses
In January 2021 HireVue, which said about 700 companies used its technology, said it would stop analyzing candidates' facial expressions in recorded video interviews. It continued to analyze intonation, speech and behavior.[34] The Electronic Privacy Information Center had filed a complaint against HireVue with the US Federal Trade Commission in 2019, and HireVue said an algorithmic audit by O'Neil Risk Consulting and Algorithmic Auditing found its software did not harbor bias; chief executive Kevin Parker said facial analysis "wasn't worth the concern."[34]
On 21 June 2022 Microsoft announced that it would retire facial analysis capabilities in Azure Face that "purport to infer emotional states" and identity attributes such as gender and age. It cited the lack of consensus on a definition of emotions, the inability to generalize the link between facial expression and emotional state, and the risk of misuse. New customers lost access that day, and existing customers had until 30 June 2023 to stop using the attributes.[9]
Regulation
European Union
Regulation (EU) 2024/1689, the AI Act, treats emotion recognition in three ways.[13]
- Prohibition. Article 5(1)(f) bans placing on the market, putting into service or using AI systems to infer the emotions of a natural person in the areas of workplace and education institutions, except where the system is intended for medical or safety reasons. It has applied since 2 February 2025.[11][37]
- High-risk classification. Annex III, point 1(c), lists AI systems intended to be used for emotion recognition as high-risk, in so far as their use is permitted under Union or national law.[38] Regulation (EU) 2026/1744, the Digital Omnibus on AI, published on 24 July 2026, moved the application date of the Annex III high-risk obligations from 2 August 2026 to 2 December 2027.[37][39]
- Transparency. Article 50(3) requires deployers of emotion recognition systems to inform people exposed to them about the system's operation, with an exception for systems permitted by law for detecting, preventing or investigating criminal offences.[40] This duty applies from 2 August 2026 and was not postponed by the Omnibus.[40][39]
The Commission published non-binding guidelines on the prohibited practices on 4 February 2025; the final version is dated 29 July 2025.[41][15] They read "workplace" broadly to include job candidates during selection and hiring, and give examples of prohibited and permitted uses.[15]
| Example in the Commission guidelines | Under Article 5(1)(f) |
|---|---|
| Call centre using webcams and voice recognition to track employees' emotions, such as anger | Prohibited |
| Call centre using voice recognition to track customers' emotions, such as anger or impatience | Not prohibited |
| Monitoring the emotional tone of hybrid work teams from voice and video calls | Prohibited |
| Emotion recognition during recruitment or a probationary period | Prohibited |
| Supermarket cameras tracking employees' emotions, such as happiness | Prohibited |
| Inferring emotions from written text to define the tone of an article | Out of scope (not biometric) |
| Inferring emotions from keystrokes, facial expressions, body postures or movements | In scope (biometric) |
Source: European Commission guidelines C(2025) 5052.[15]
Outside workplaces and education, the guidelines say emotion recognition systems are not banned by Article 5(1)(f) but are high-risk, and that some uses may still fall under the Act's bans on harmful manipulation and exploitation or under other EU law such as data protection rules.[15] The facial recognition and GDPR articles cover the wider biometric rules.
United Kingdom
In October 2022 the Information Commissioner's Office warned organisations against using biometric technologies for emotion analysis, said it had yet to see any emotion AI technology develop in a way that satisfies data protection requirements, and said it planned to publish guidance on biometric technologies in spring 2023.[10]
United States
California Assembly Bill 1883, by Assemblymember Isaac Bryan, was signed by Governor Gavin Newsom on 30 September 2026 and adds a part on workplace surveillance tools to the Labor Code.[12][42] It takes effect on 1 January 2027.[43] It bars employers, including public employers, from using a workplace surveillance tool that uses artificial intelligence on employees to recognize or make inferences or predictions about an individual's emotional state, or to collect neural data. Tools used to ensure safety are not prohibited, and some federally required aerospace and defense work is exempt.[12] The Labor Commissioner or a public prosecutor can enforce it, with a penalty of up to $500 per violation.[12]
Continue reading
- EU AI ActThe EU AI Act is the common name for Regulation (EU) 2024/1689, a directly applicable European Union law that sets horizontal rules for artificial intelligence.Mentioned 3 times above
- Sentiment AnalysisSentiment analysis, also called opinion mining or emotion AI, is the use of natural language processing to automatically determine the sentiment, emotion, or opinion expressed in a piece of text, typically…Mentioned twice above
- Facial RecognitionFacial recognition is a biometric technology that identifies or verifies the identity of an individual by analyzing patterns in a digital image or video frame of the person's face.Listed in See also
- Speech RecognitionSpeech recognition, usually called automatic speech recognition (ASR), is the computational task of converting a spoken-language signal into a sequence of written symbols.Listed in See also
- Algorithmic biasAlgorithmic bias is the tendency of a computer system to produce systematic, repeatable errors that advantage some groups of people over others.Listed in See also
- AI in EducationAI in education refers to the use of artificial intelligence technologies to support teaching, learning, assessment, and administrative processes across educational settings.Listed in See also
More to read
Topics:AI EthicsAI Policy & RegulationComputer VisionSpeech & Audio AI
References
- ^1 ^2 ^3Mohammad, Saif M. "Ethics Sheet for Automatic Emotion Recognition and Sentiment Analysis". Computational Linguistics, vol. 48, no. 2, pp. 239-278, 2022. doi.org/...coli_a_00433
- ^1 ^2 ^3Crawford, Kate; Dobbe, Roel; Dryer, Theodora; et al. "AI Now 2019 Report". AI Now Institute, December 2019. ainowinstitute.org/...ai-now-2019-report-2
- ^1 ^2 ^3Picard, R. W. "Affective Computing". M.I.T. Media Laboratory Perceptual Computing Section Technical Report No. 321, revised 26 November 1995. vismod.media.mit.edu/...TR-321.pdf
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Goodfellow, Ian J.; Erhan, Dumitru; Carrier, Pierre Luc; Courville, Aaron; Mirza, Mehdi; et al. "Challenges in Representation Learning: A report on three machine learning contests". arXiv:1307.0414, 1 July 2013. arxiv.org/...1307.0414
- ^1 ^2 ^3 ^4Yang, Shu-wen; Chi, Po-Han; Chuang, Yung-Sung; et al. "SUPERB: Speech processing Universal PERformance Benchmark". arXiv:2105.01051, 3 May 2021. arxiv.org/...2105.01051
- ^1 ^2 ^3 ^4Demszky, Dorottya; Movshovitz-Attias, Dana; Ko, Jeongwoo; Cowen, Alan; Nemade, Gaurav; Ravi, Sujith. "GoEmotions: A Dataset of Fine-Grained Emotions". arXiv:2005.00547, 1 May 2020. arxiv.org/...2005.00547
- ^1 ^2 ^3Koelstra, S.; Muhl, C.; Soleymani, M.; Lee, Jong-Seok; Yazdani, A.; Ebrahimi, T.; Pun, T.; Nijholt, A.; Patras, I. "DEAP: A Database for Emotion Analysis Using Physiological Signals". IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18-31, January 2012. doi.org/...T-AFFC.2011.15
- ^1 ^2 ^3 ^4 ^5Barrett, Lisa Feldman; Adolphs, Ralph; Marsella, Stacy; Martinez, Aleix M.; Pollak, Seth D. "Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements". Psychological Science in the Public Interest, vol. 20, no. 1, pp. 1-68, July 2019. doi.org/...1529100619832930
- ^1 ^2Bird, Sarah. "Responsible AI investments and safeguards for facial recognition". Microsoft Azure Blog, 21 June 2022. azure.microsoft.com/...ards-for-facial-recognition
- ^1 ^2 ^3Klovig Skelton, Sebastian. "ICO warns against using biometrics for 'emotional analysis'". Computer Weekly, 26 October 2022. computerweekly.com/...trics-for-emotional-analysis
- ^1 ^2Future of Life Institute. "Article 5: Prohibited AI Practices". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689 as amended), accessed 18 November 2026. artificialintelligenceact.eu/article
- ^1 ^2 ^3 ^4California Legislature. "AB-1883 Workplace surveillance tools (2025-2026)". California Legislative Information. leginfo.legislature.ca.gov/...billNavClient.xhtml
- ^1 ^2Future of Life Institute. "Article 3: Definitions". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), accessed 18 November 2026. artificialintelligenceact.eu/article
- ^Future of Life Institute. "Recital 18". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), accessed 18 November 2026. artificialintelligenceact.eu/...18
- ^1 ^2 ^3 ^4 ^5 ^6European Commission. "Commission Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act)". Communication C(2025) 5052 final, Brussels, 29 July 2025. ec.europa.eu/...112367
- ^1 ^2 ^3 ^4Mohammad, Saif; Bravo-Marquez, Felipe; Salameh, Mohammad; Kiritchenko, Svetlana. "SemEval-2018 Task 1: Affect in Tweets". Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pp. 1-17, 2018. doi.org/...S18-1001
- ^Ekman, Paul; Friesen, Wallace V. "Constants across cultures in the face and emotion". Journal of Personality and Social Psychology, vol. 17, no. 2, pp. 124-129, 1971. doi.org/...h0030377
- ^1 ^2Paul Ekman Group. "Facial Action Coding System". Paul Ekman Group, accessed 18 November 2026. paulekman.com/facial-action-coding-system
- ^1 ^2 ^3 ^4 ^5 ^6Lucey, Patrick; Cohn, Jeffrey F.; Kanade, Takeo; Saragih, Jason; Ambadar, Zara; Matthews, Iain. "The Extended Cohn-Kanade Dataset (CK+): A complete dataset for action unit and emotion-specified expression". 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops, pp. 94-101, June 2010. doi.org/...CVPRW.2010.5543262
- ^Russell, James A. "A circumplex model of affect". Journal of Personality and Social Psychology, vol. 39, no. 6, pp. 1161-1178, 1980. doi.org/...h0077714
- ^Belaref, Amdjed; Sadok, Samir; Noumir, Zineb; Seguier, Renaud. "Data-Driven Decoding of Russell's Circumplex Model of Affect". arXiv:2606.16843, 15 June 2026. arxiv.org/...2606.16843
- ^1 ^2 ^3 ^4 ^5 ^6Mollahosseini, Ali; Hasani, Behzad; Mahoor, Mohammad H. "AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild". IEEE Transactions on Affective Computing, vol. 10, no. 1, pp. 18-31, 2019 (first published online 2017). doi.org/...TAFFC.2017.2740923
- ^1 ^2 ^3 ^4Signal Analysis and Interpretation Laboratory (SAIL), University of Southern California. "The Interactive Emotional Dyadic Motion Capture (IEMOCAP) Database". Accessed 18 November 2026. sail.usc.edu/iemocap
- ^Picard, Rosalind W. "Affective Computing". The MIT Press, 1997. doi.org/...1140.001.0001
- ^1 ^2Picard, R. W.; Vyzas, E.; Healey, J. "Toward machine emotional intelligence: analysis of affective physiological state". IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 10, pp. 1175-1191, 2001. doi.org/...34.954607
- ^Picard, Rosalind W. "Affective Computing: From Laughter to IEEE". IEEE Transactions on Affective Computing, vol. 1, no. 1, pp. 11-17, January 2010. doi.org/...T-AFFC.2010.10
- ^1 ^2Schuller, Björn; Steidl, Stefan; Batliner, Anton. "The INTERSPEECH 2009 emotion challenge". Interspeech 2009, pp. 312-315, 2009. doi.org/...Interspeech.2009-103
- ^1 ^2Chen, Sanyuan; Wang, Chengyi; Chen, Zhengyang; et al. "WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing". arXiv:2110.13900, 26 October 2021. arxiv.org/...2110.13900
- ^Schuller, Björn W. "Speech emotion recognition: two decades in a nutshell, benchmarks, and ongoing trends". Communications of the ACM, vol. 61, no. 5, pp. 90-99, 2018. doi.org/...3129340
- ^Eyben, Florian; Scherer, Klaus R.; Schuller, Bjorn W.; Sundberg, Johan; Andre, Elisabeth; Busso, Carlos; Devillers, Laurence Y.; Epps, Julien; et al. "The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing". IEEE Transactions on Affective Computing, vol. 7, no. 2, pp. 190-202, April 2016. doi.org/...TAFFC.2015.2457417
- ^1 ^2 ^3Livingstone, Steven R.; Russo, Frank A. "The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English". PLOS ONE, vol. 13, no. 5, e0196391, 16 May 2018. doi.org/...journal.pone.0196391
- ^1 ^2 ^3Poria, Soujanya; Hazarika, Devamanyu; Majumder, Navonil; Naik, Gautam; Cambria, Erik; Mihalcea, Rada. "MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations". Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 527-536, 2019. doi.org/...P19-1050
- ^Smart Eye. "Smart Eye Acquires Affectiva to Solidify Stronghold on Interior Sensing Market". Smart Eye, 25 May 2021. smarteye.se/...ronghold-on-interior-sensing-market
- ^1 ^2 ^3Knight, Will. "Job Screening Service Halts Facial Analysis of Applicants". Wired, 12 January 2021. wired.com/...vice-halts-facial-analysis-applicants
- ^1 ^2Cowen, Alan S.; Keltner, Dacher; Schroff, Florian; Jou, Brendan; Adam, Hartwig; Prasad, Gautam. "Sixteen facial expressions occur in similar contexts worldwide". Nature, vol. 589, pp. 251-257, 2021 (published online 16 December 2020). doi.org/...s41586-020-3037-7
- ^1 ^2Future of Life Institute. "Recital 44". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), accessed 18 November 2026. artificialintelligenceact.eu/...44
- ^1 ^2Future of Life Institute. "Article 113: Entry into Force and Application". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689 as amended), accessed 18 November 2026. artificialintelligenceact.eu/...113
- ^Future of Life Institute. "Annex III: High-Risk AI Systems Referred to in Article 6(2)". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), accessed 18 November 2026. artificialintelligenceact.eu/annex
- ^1 ^2Pearce, Sarah; Nietsch, Thomas (K&L Gates). "EU Digital Omnibus on AI Enters Into Force". The National Law Review, 31 July 2026. natlawreview.com/...igital-omnibus-ai-enters-force
- ^1 ^2Future of Life Institute. "Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems". EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), accessed 18 November 2026. artificialintelligenceact.eu/...50
- ^European Commission. "Commission publishes the Guidelines on prohibited artificial intelligence (AI) practices, as defined by the AI Act". Shaping Europe's digital future, 4 February 2025 (updated 31 July 2025). digital-strategy.ec.europa.eu/...es-defined-ai-act
- ^1 ^2Office of Governor Gavin Newsom. "California's nation-leading AI framework just got stronger, Governor Newsom signs more first-in-the-nation worker protections and more". Governor of California, 30 September 2026. gov.ca.gov/...e-nation-worker-protections-and-more
- ^1 ^2Hunton Andrews Kurth. "California Enacts New Restrictions on AI-Based Employee Monitoring". Privacy & Cybersecurity Law Blog, 6 October 2026. hunton.com/...ions-on-ai-based-employee-monitoring
- ^Busso, Carlos; Bulut, Murtaza; Lee, Chi-Chun; Kazemzadeh, Abe; Mower, Emily; Kim, Samuel; Chang, Jeannette N.; Lee, Sungbok; Narayanan, Shrikanth S. "IEMOCAP: interactive emotional dyadic motion capture database". Language Resources and Evaluation, vol. 42, no. 4, pp. 335-359, 2008. doi.org/...s10579-008-9076-6
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
2 revisions · v2 · 4,416 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 96 claims checked against 46 cited and independent sources; 1 material and 4 minor corrections applied before publication.
Cite this page: AI Wiki. "Emotion recognition." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/emotion_recognition