What is running, what it was trained on, and what has not been measured yet.
No evaluation results exist yet. This is a proposal-stage research build. Not yet evaluated on local data.
| Engine | State | Training data | Evaluated on local data | Known limits |
|---|---|---|---|---|
| Text — published model Shahwaiz75/menzync-emotion-model (XLM-RoBERTa base, 7 labels) | Not connected | Unknown — the original training data and notebooks are not available, so the reported accuracy cannot be reproduced. | No | No model card or licence published yet. Roman Urdu spelling variety and sarcasm are known weak points for text models. |
| Text — rule-based baseline menzync-lexicon-baseline-2 | Fallback active | None. A hand-written Roman Urdu and English cue list with negation handling. | No | Not a learned model. Reports relative strength, never probability, and is never averaged with other signals. |
| Voice superb/wav2vec2-base-superb-er (planned) | Not connected | IEMOCAP — acted English speech, four classes (neutral, happy, angry, sad). | No | Spoken Urdu is outside its training data. Acted emotion differs from everyday speech. |
| Facial expression @vladmandic/face-api@1.7.15 face_expression_model + tiny_face_detector | On your device | Third-party pretrained weights; the publisher does not document the training data. | No | Expression is not emotion. Lighting, pose, culture and context change the estimate. Runs on the device; no identity model is loaded. |
Weighted late fusion over the signals that produced a usable reading. A missing signal never becomes "neutral". Probabilities are only combined with probabilities; the baseline’s relative scores are shown but never averaged in. When channels point in different directions, the result says so, and a tie produces no reading at all.
A confidence figure is only meaningful when a model has been calibrated against held-out data from the people who use it. That has not happened yet, so Menzync shows the reading, the engine that produced it, and what it could not read — and leaves the final word to you.
Status checked live: Sun, 01 Nov 2026 19:26:57 GMT · /api/v1/capabilities