Read the full spectrum of human emotion from a single image.
Most emotion detection APIs tell you "happy" or "sad." Our facial expression analysis API scores 48 distinct emotions at once from Joy and Anger to Nostalgia, Contemplation and Entrancement in real time. Turn on your webcam and watch the page react to your emotional fingerprint.
No signup to try · Runs in your browser · Backed by IIM Ahmedabad & 100X.VC
Live Emotion Detection
Enable your camera to see detected emotions in real-time.
Emotion has a structure. We mapped it.
Every reading you saw above is assembled from 48 independent dimensions scored in parallel, and the strand below carries every one of them. Move your cursor to turn it.
A face is never just one feeling. Single-label detection throws away the joy, the warmth, the hint of nostalgia you read instantly. We built a model that doesn't.
Emotions are simultaneous
A person can feel Joy and Anxiety, Admiration and Awkwardness. We score every emotion independently no forced choice.
The signal is subtle
The difference between Contempt and Concentration is a few facial muscles. Coarse labels miss it; a 48-dimension readout captures it.
The real world is messy
Odd lighting, angles, occlusions even cartoons and paintings. The model is built to generalize beyond clean lab photos.
Not a label. A measurement.
For every image, the API returns an independent probability score for each of 48 emotions a full distribution, not a winner. A 0.45 on Nostalgia tells a different story than a 0.02. That distribution is the insight.
Forcing one emotion discards information. The probability spread lets you set your own thresholds, track blends, and feed nuanced signal into downstream models or analytics.
Readout · "laughing at a joke"
Expression is a signal, not a snapshot.
A single frame tells you what a face looks like. A stream of frames tells you what just happened the instant a punchline lands, a patient winces, or a viewer checks out. The streaming endpoint scores every frame, so you get 48 traces over time instead of one number. Watch a twelve-second clip play out below or drag the playhead yourself.
Live trace · "hearing a joke" · 4 of 48 dims shown
Why this matters: every downstream use case ad pre-testing, adaptive tutoring, clinical monitoring, empathic agents lives in the deltas. The API gives you the raw traces; your product decides what a spike means.
All 48 emotions, on one spectrum.
Every dimension the model measures, arranged from calm and low-arousal to intense and high-arousal. Hover any chip for a definition.
How a photo becomes 48 numbers.
One request travels through four stages. Each stage exists for a reason and the whole trip is fast enough to run on a live webcam stream.
Find the face
A detector locates every face in the frame off-angle, partially lit, or half-turned. No face, no charge: the request short-circuits before inference.
Align & normalize
Landmarks anchor the crop so tilt, distance and camera angle stop mattering. The model compares expressions, never framing.
Encode holistically
A transformer vision backbone reads the whole face at once. It learns that tension around the eyes means one thing next to a relaxed mouth and another next to a clenched jaw relations, not regions.
Score all 48 heads
Forty-eight independent output heads fire in parallel one per emotion dimension. No softmax forcing a winner: every score stands on its own, so blends survive.
Distinct human faces
A proprietary training corpus deliberately weighted toward rare expressions Entrancement, Empathic Pain, Aesthetic Appreciation so the model learns them instead of ignoring them.
Independent probabilities
Each head is trained and calibrated separately. A 0.45 on Nostalgia means the same thing today as it did last month scores you can threshold and trend against.
Live-stream ready
The full pipeline is optimized for streaming the same model that scores a batch of stills keeps pace with a webcam feed over WebSocket.
Stateless by default
Frames are scored and released. Nothing about the pipeline requires retaining imagery retention is your policy decision, not our architecture's.
Three ways in.
Upload
Drop in any local image JPG or PNG straight from your device.
Paste a URL
Point the model at any hosted image and get scores back instantly.
Live webcam
Stream from your camera and watch your emotional fingerprint update in real time.
Built for the corners where generic models fail.
Foundation models can do zero-shot emotion recognition but they flatten the truth. Our 48-dimension granularity, real-time speed and regulatory-ready architecture unlock the verticals where "good enough" isn't an option.
Non-verbal pain assessment
The strongest clinical signal for pain in those who can't self-report, including neonates, dementia patients, and the critically ill. Continuous monitoring replaces intermittent, subjective spot-checks.
Falls under the EU AI Act medical exceptionEldercare & proactive companions
Emotion-aware virtual coaches for older adults, with a companion that initiates contact when it detects stress or anxiety instead of waiting to be asked.
Youth mental-health screening
Screening support for adolescents in telehealth and clinical settings, with guardian consent only and never in schools.
Sensitive · consent + clinical oversight requiredSignals that matter here
Pain and distress dimensions dominate, providing the signals clinicians need when a patient can't self-report.
Because general vision-language models can't touch them whether it's the 48-dimension granularity a clinical pain scale needs, regulatory clearance for a medical device, or custom models for veterinary care. This is where we win.
A few lines to your first prediction.
Send an image, get 48 scores back. REST for batches, streaming for live webcam and video.
import requests resp = requests.post( "https://api.sharedpro.co/v1/expression/face", headers={"Authorization": "Bearer YOUR_API_KEY"}, json={"image_url": "https://example.com/face.jpg"}, ) # -> {"emotions": [{"name": "Joy", "score": 0.86}, ...]} # 48 dims for e in resp.json()["emotions"][:5]: print(f'{e["name"]:<14}{e["score"]:.2f}')
What this measures and what it doesn't.
These scores reflect how an average human observer would likely describe an expression. They are not a direct readout of what someone is feeling emotional experience is private, multimodal and context-dependent. Use the API to augment human judgment, not replace it, and always with the consent of the people being analyzed especially in sensitive verticals like clinical pain assessment or youth mental health.
Built on Hume's Expression Measurement API? It sunset on June 14, 2026.
If your product depended on Hume AI's face model, Sharedpro is the closest drop-in for facial expression analysis the same 48-dimension emotion taxonomy, independent per-dimension scores, REST for batches plus streaming for live video, and a free tier to validate the swap before you commit.
Start free.
- Unlimited browser demo
- Webcam, upload & URL
- No signup required
- Full REST + streaming API
- 48-emotion output
- Free monthly call allowance
- Docs & SDKs
- Volume pricing & SLA
- On-platform integration
- Clinical / custom-model builds
- Priority support
Questions, answered.
What does the API return?
For each image, an independent confidence score for all 48 emotion dimensions a full probability distribution, not a single label. You decide how to rank, threshold, or combine them.
What inputs are supported?
JPG and PNG images, supplied as a file upload, a hosted URL, or a live webcam stream. REST handles single images and batches; the streaming endpoint handles real-time webcam and video.
Is this a readout of how someone really feels?
No. The scores reflect how an average observer would describe an expression, not a person's private emotional experience. See Responsible use augment human judgment, with consent.
How fast is it?
Inference is optimized for real-time use fast enough to score a live webcam frame-by-frame, not just process batches offline.
Is this an alternative to Hume AI's Expression Measurement API?
Yes. Hume AI sunset its Expression Measurement API on June 14, 2026, ending new jobs and API access. Sharedpro's face model outputs the same 48-dimension emotion taxonomy with independent per-dimension scores, over REST (batch) and streaming (real-time), with a free tier to test the migration. See migrating from Hume.
How do you handle privacy?
Process only media you have the right to analyze, with the consent of the people in it. Use signed/expiring URLs for hosted files, and avoid storing raw imagery longer than your use case requires.
See your own 48 emotions.
Turn on your camera, or grab an API key and ship emotional intelligence into your product this week.