Emotion Recognition APIs Compared (2026)
Products described as emotion recognition can return very different kinds of facial-expression output. Compare published capabilities, delivery models, and access conditions before deciding whether a tool fits your review task.
Published and updated November 1, 2026.
Published capabilities as of November 1, 2026
| Provider | What it returns | Delivery | Output format | Status and access | Checked | Source |
|---|
| AWS Rekognition | Emotion types with confidence | Cloud API | Named categories and confidence | Not a determination of internal emotional state | checked Nov 1, 2026 | Emotion API reference |
| Google Cloud Vision | Face boxes, landmarks, expression likelihoods, quality attributes | Cloud API | VERY_UNLIKELY to VERY_LIKELY ratings | No recognition of specific individuals | checked Nov 1, 2026 | Detect faces |
| Microsoft Azure Face | Emotion-related attributes retired | Cloud API | No current emotion-attribute offering in the cited notice | Unavailable to new customers June 21, 2022; existing customers after June 30, 2023 | checked Nov 1, 2026 | Responsible AI investments and safeguards for facial recognition |
| Hume AI Expression Measurement | Speech and face expression scores | Upload and realtime WebSocket interfaces; Python and Node.js SDKs | Probabilities from 0 to 1 | Limited release; contact Hume for access | checked Nov 1, 2026 | Expression Measurement API overview |
| MorphCast | Vendor-described 100+ affective signals | Client-side JavaScript SDK, browser or app | Signals described by the SDK; no format assumed here | Per vendor, images and video are not sent to a server | checked Nov 1, 2026 | MorphCast SDK |
| Mental Edge | Time-coded expression and movement signals for human review | Developer API, asynchronous jobs | Frame-by-frame category scores; Queued, Processing, Completed, Failed job states | See developer pricing; no prices compared here | checked Nov 1, 2026 | Mental Edge developer pricing |
AWS source, as of November 1, 2026
Google source, as of November 1, 2026
Microsoft source, as of November 1, 2026
Hume source, as of November 1, 2026
MorphCast source, as of November 1, 2026
Mental Edge developer pricing
Start with the output you need
A named expression category, a landmark coordinate, and a movement annotation answer different questions. Decide which one your project needs before comparing provider labels. A convenient product name should not replace a written description of the intended task.
For a recorded-video review, ask how reviewers will return to a specific source moment. For a landmark task, separate geometry evaluation from any interpretation attached to the face. For an expression task, define the annotation rules and uncertainty categories before selecting a service.
This comparison summarizes only the published facts identified in the Sources list. It is not a speed, accuracy, privacy, or value ranking. Unknown details remain questions for the provider rather than assumptions that make a table look complete.
AWS Rekognition: confidence is not inner-state certainty
As of November 1, 2026, AWS lists HAPPY, SAD, ANGRY, CONFUSED, DISGUSTED, SURPRISED, CALM, FEAR, and UNKNOWN. Its Emotion reference describes a confidence value alongside the type. AWS explicitly states that the result is not a determination of the person's internal emotional state.
Keep that caveat attached when presenting category counts or reviewing individual examples. A confidence value should not be relabeled as confidence that someone genuinely feels the named emotion. Compare the service's categories with your codebook rather than silently treating them as equivalent.
If your task concerns visible movement, decide how the named categories will be inspected against source footage. Record disagreements instead of dismissing them because a number appears confident. The published category list does not establish suitability for employment, clinical, or educational decisions.
Google Cloud Vision: likelihood ratings and face attributes
As of November 1, 2026, Google describes bounding boxes, landmarks, and expression likelihoods in FACE_DETECTION. The expression labels cover joy, sorrow, anger, and surprise, with ratings from VERY_UNLIKELY to VERY_LIKELY. Its face attributes also include headwear, blur, and under-exposure.
Google's cited documentation says facial recognition of specific individuals is not supported. Do not confuse detecting a face with identifying the person shown. That distinction matters when planning consent, reviewing sensitive recordings, or explaining the system to participants.
Likelihood categories are not automatically interchangeable with another provider's numerical scores. Preserve the original output terminology in your evaluation record. If you create a mapping for comparison, document it and explain what information that mapping loses.
Microsoft Azure: account for retired attributes
Microsoft announced the retirement of Face API emotion-related attributes on June 21, 2022. They became unavailable to new customers that day and to existing customers after June 30, 2023. The notice also covers attributes including age, gender, and smile.
Microsoft cites privacy, disagreement about the definition of emotions, and problems generalizing expression-to-state relationships. Describe these as Microsoft's stated reasons rather than inventing another explanation for the decision. The retirement notice is not a current emotion-analysis capability to include in a shortlist.
If replacing an older integration, first inventory which outputs the previous workflow actually consumed. Do not assume a similarly named label from another system preserves the original meaning. A migration needs a documented review question, not simply a replacement field with a familiar name.
Hume AI: check current access and model scope
As of November 1, 2026, Hume describes Expression Measurement as limited release, with access requested through Hume. The documentation describes expression in speech and faces through upload and realtime WebSocket interfaces. It lists probabilities from zero to one and Python and Node.js SDKs.
As of November 1, 2026, its video description covers forty-eight expression names and twenty-seven description names. Scores concern the largest faces, with two selected by default in the cited overview. Hume's marketing page focuses on emotionally intelligent voice AI.
Do not infer that the marketing emphasis guarantees or removes a documented expression capability. Ask about access before designing a workflow around a limited-release service. Record which modality and face-selection behavior your evaluation actually uses rather than comparing unlike inputs.
MorphCast: a different delivery model
As of November 1, 2026, MorphCast describes a client-side JavaScript engine running in browsers or apps. It says video and pictures are not sent to a server. The vendor describes a library under one megabyte and more than one hundred affective signals.
These are MorphCast's published descriptions, not independent measurements from this article. A client-side SDK and a cloud service create different integration questions. Decide where review happens and what material your team needs to retain before choosing between delivery models.
Do not turn a stated delivery model into an absolute privacy guarantee. Your own application still needs a clear purpose, appropriate consent, and controlled handling of recordings. Check the implementation you plan to build rather than assuming the library description covers the entire application.
Where Mental Edge fits
Mental Edge surfaces observable expression and movement signals from recorded video as time-coded outputs. It scores expressions frame by frame across Enjoyment, Sadness, Anger, Fear, Surprise, Disgust, and Contempt. Those patterns may indicate moments for contextual review, not verified feelings, intent, or truthfulness.
The developer API includes face detection, embeddings, a face gallery, and asynchronous facial-expression analysis. Published job states are Queued, Processing, Completed, and Failed. Image work is billed per image, video per minute, and failed jobs are not billed.
Use the developer pricing page for current commercial details rather than relying on a copied price table. Mental Edge is not for hiring, promotion, diagnosis, deception detection, threat assessment, student scoring, or surveillance. No capability described here overrides those boundaries.
Evaluate with held-out clips and independent review
Write a small evaluation plan before reviewing attractive examples from any provider. Define the input material, consent scope, codebook, quality exclusions, and outputs being compared. Keep some clips separate from the material used to refine your rules.
Have independent reviewers inspect the same source moments using the agreed definitions. Preserve uncertainty and disagreement instead of forcing every example into a label. Compare each provider's outputs with that record while keeping category mappings explicit.
Our provider-migration workflow offers a planning structure for this process. The facial-analysis benchmark page should be read as the protocol or evidence it actually presents. A proposed evaluation does not establish results, and this comparison reports no comparative performance findings.
Why providers caveat expression output
Barrett and colleagues' 2019 review describes variation across cultures, situations, and individuals. The same facial configuration can express more than one emotion or none. AWS's internal-state caveat and Microsoft's stated concerns illustrate why output interpretation needs careful boundaries.
Avoid presenting a provider's confidence or probability as a participant's self-reported experience. Return to the source moment and any context the review was permitted to use. If the question requires an internal state, acknowledge what the available evidence cannot answer.
Regulation and selection checklist
EU AI Act Article 5(1)(f) prohibits workplace and educational emotion inference, except for medical or safety reasons. The prohibitions apply from February 2, 2025. Talk to counsel about your use case. This comparison is not legal advice.
Before choosing, confirm output definitions, current access, delivery model, source-review needs, and applicable terms. Ask whether your intended purpose is allowed before evaluating technical convenience. Document unanswered questions and reject uses that turn uncertain expression estimates into high-stakes judgments.