Spell it silently.
OurASL decodes one-handed ASL fingerspelling from the motion sensors in Oura rings — so you can type into smart glasses without speaking a word, and without anything to type on.
Wear. Spell. Decode. Read.
Wear
Put on an Oura ring — possibly the one you already sleep in. Its motion sensors stream about fifty readings a second over Bluetooth.
Spell
Fingerspell one-handed ASL, or draw letters in the air with a single ring. No pausing between letters — write the way you'd speak.
Decode
A temporal neural network reads the motion stream and resolves it into letters; a language model turns letters into words, fixing the ambiguous ones.
Read
Words land wherever you point them — smart glasses, a phone, a terminal. Nothing spoken. Nothing typed. Nothing overheard.
Handshapes, or letters drawn in the air
Air trace works today · one ring
Draw capital letters in the air and the ring's accelerometer follows the stroke. Each letter is a motion trajectory — no handshapes to learn, and a single ring is enough. This is the mode you can train and test right now.
ASL fingerspelling in training
The destination: the real one-handed ASL alphabet, 26 letters plus space, backspace, and enter. The vocabulary and capture trainer are built; the multi-ring capture that makes it fluent is the next phase.
An open pipeline, end to end
Every stage is open source and tested — from the Bluetooth frames to the decoded text. No cloud, no account, no audio: the model runs on your machine and the words never leave it.
- Real hardware, pure Python. A bleak-based Bluetooth bridge streams a physical Oura Ring 4 — scan, pair, and stream from the CLI.
- 48 Hz motion frames, resampled onto a common clock so multiple rings can join later without touching the model.
- Temporal CNN + CTC — the same decoding family used for speech and handwriting, pointed at your knuckles.
- A browser capture trainer with hand diagrams and animated stroke guides, so building your own dataset takes an afternoon.
- No ring required to start — a physics simulator stands in for hardware, and the whole test suite runs against it.
Voice assistants taught computers to listen — and made everyone nearby listen too. OurASL starts from a different premise: the most private input device you own is your hand. Fingerspelling is a complete alphabet the Deaf community has refined for centuries, and a ring is a computer that already lives on your finger.
Put them together and you get text entry that is silent, one-handed, and eyes-free — in a meeting, on a train, in the dark, with a coffee in your other hand. An input method is a language, and languages shouldn't have owners.
What works, and what's next
OurASL is a research project. The full pipeline runs end to end today — driven by both a physics simulator and a real Oura Ring 4 over Bluetooth — but recognition is still lab-grade and the vocabulary is still growing. Here's exactly where it stands:
- done End-to-end pipeline: simulator → features → TCN+CTC → decoded text, fully tested
- done Real Oura Ring 4 over Bluetooth — pure-Python scan, pair, and live streaming
- done Browser capture trainer for ASL handshapes and air-traced letters
- now Training on real captures — teaching the decoder a human's actual alphabet
- next Multi-ring capture for fluent ASL handshapes
- next Language-model beam decoder — letters in, words out
Teach it your alphabet
If you have an Oura ring and some patience, you can log your first training captures this afternoon — and every alphabet it learns makes the decoder better for everyone.