AIHumanity
Emotion Models

Pre-trained Emotion Recognition Across Face, Voice, and Text

Production-ready emotion models for face, voice, and text — benchmarked, calibrated, and ready to drop into your stack. ONNX, TFLite, and server-side variants.

Get model accessTalk to us
3 modalities
face · voice · text
ONNX + TFLite
edge + server
Benchmarked
open evaluation set

Three modalities. One coherent label space.

Designed to be fused — see also [[fusion]] for the joint model.

Face

Facial Expression Recognition

Lightweight FER model trained on a curated multi-source dataset. Returns categorical, dimensional (valence/arousal), and AU outputs from a single forward pass.

  • ›Mobile-friendly ONNX + TFLite
  • ›AU + categorical + valence/arousal
  • ›Robust across skin tones and lighting
Voice

Speech Emotion Recognition

Acoustic + prosodic features feed a transformer head that outputs the same emotion label space as the FER model — built for downstream fusion.

  • ›Low-latency streaming inference
  • ›Speaker-invariant features
  • ›Coherent with face label space
Text

Text Emotion Recognition

Distilled transformer for short conversational text. Calibrated probabilities, not just argmax — built for use inside a fusion pipeline.

  • ›Calibrated probabilities
  • ›Conversational + formal modes
  • ›Tiny variant for on-device chat

Where it fits

🎮

Games

NPC reactions tied to real player emotion.

🤖

Assistants

Emotion-aware response generation.

📞

Contact centres

Real-time sentiment for live calls.

📱

Consumer apps

On-device emotion read for mobile features.

Build with emotion that actually generalises

Request developer access — model cards, benchmarks, and integration guides included.

Get developer access →
Talk to David