Skip to main content

Keyboard shortcuts

In-browser · privacy-first

Read sign language in your browser

100% in-browser inference · webcam frames never leave your device · MobileNetV2, 59.8% honest cross-dataset accuracy (A–Y) vs 96.9% same-dataset.

Real-world webcam accuracy is lower than the benchmark — see How it works.

Live demo

Point your webcam at an ASL alphabet sign and watch the prediction update in real time. Your hand is detected and cropped, then classified — all on-device. Frames never leave your browser.

Loading live demo

Upload an image

Prefer not to use a webcam? Drop in a photo of a hand sign — or click an example — and get a classification with top-5 probabilities.

Loading upload panel

See what the model sees

Freeze a frame and step through the inference pipeline: hand detection, crop, tensor normalization, and the final prediction. Drag the temperature slider to watch the probability distribution reshape in real time.

Loading explainer

Metrics

Every number below is produced by reproducible code in the repo and measured on the held-out test set — nothing is hardcoded.

Loading metrics

How it works

Hands are detected with MediaPipe, cropped, resized, and ImageNet-normalized, then run through a MobileNetV2 ONNX model with onnxruntime-web — entirely in the browser. Real-world accuracy is lower than the benchmark: the honest cross-dataset number is 59.8% (A–Y) vs the 96.9% same-dataset benchmark. Read the full story and model card on the About page.

Read the project story & model card →