Keyboard shortcuts: Alt + Shift + S to speak, E to edit, X to stop. Shortcuts pause while typing.
Build a sentence
Sign one word, then pause. Add the match to your draft, or turn on automatic adding below. Review the draft before speaking.
Automatic adding can include incorrect words. Review your draft before selecting Speak message.
Typed words & sentence tools
Typed words:
Try “I AM HUNGRY.” or “I WANT WATER, PLEASE.” Typed words are not camera detections. This builds reviewed text from separate signs; it does not translate continuous ASL or its grammar. Use Edit message for corrections.
Review before saving or speaking. Saved text is stored on this server; camera frames stay on your device.
✓
Hold each sign steadyLetters are added after a stable detection.
↻
Move between signsLower your hand briefly before repeating a letter.
Saved history is private to this browser. Clearing cookies removes your access.
Connecting to saved messages…
No saved sessions yet.
0Saved sessions
0Detections
0%Average agreement · not probability
0 msAverage letter hold time
Measured recognition results
Five-word server preview · DTW · updated 18 October 2026. Isolated ASL Citizen clips with separate training, validation, and test signers. Tracking failures count as misses.
Current five-word model: held-out clip results
Word
Correct
Missed
Wrong word
NAME
8 / 11
3
0
HOME
8 / 11
3
0
HAVE
11 / 11
0
0
WANT
15 / 22
7
0
LOVE
9 / 11
2
0
51 / 66 correct; 15 misses. Unsupported signs: 3 false matches in 48 clips. The model did not pass the original zero-false-match release criterion and remains an optional preview.
These results do not measure live webcam accuracy, automatic movement boundaries, or ordinary hand movements. No calibrated confidence percentage is available.
Model identity and other evaluations
Deployed model SHA-256: ecb9c8e061876c9ca50af4b9f5faab55d78a530798ac653798f9e95c37fefb67
These are user ratings, not measured model accuracy. Letter tests · Motion tests. Submit the study’s 5-point evaluation for accuracy, usability, and performance. No account identity is attached to feedback. Avoid personal information in feedback.
0 responsesAccuracy —Usability —Performance —
Recorded recognition checks
Whole-clip checks on different recorded subsets. These are not live-camera confidence or continuous-sentence accuracy. Default misses include tracking failures; unsupported-sign counts use tracked negative clips. User feedback is reported separately.
Model / subset
Correct
Wrong word
Missed
False matches on other signs
Quick reference
Learn supported signs
Numbers 0–9
ASL learning · experimental recognition
Say it with a sign.
Select Words in motion and start the camera. Automatic words detects movement and suggests a word after you pause. Review the result at the camera. You can also use Record a word for a timed capture. Personal training is optional.
Watch & practice
See the movement.
Choose a word, watch the full movement, then try it at the camera.
HELLO · Automatic capture: start the camera, sign once, then pause. Review the match.
Loading recorded test accuracy…
Watch the hand shape, direction, and position near the face or body. Keep your hands and upper body visible when practicing. Examples show one signing variant; recognition may still miss a correct sign.
Examples: PopSign ASL v1.0, Georgia Tech & DPAN, CC BY 4.0. Resized and compressed. The seven core-word clips come from the ASL Citizen research dataset (Microsoft Research) and keep its dataset terms. Full credits.
Explore 201 words & phrases
ASL-Citizen research dataset: Microsoft documents about 2,731 isolated ASL signs and 84,000 recordings. Its dataset terms do not permit republishing the complete data catalog here — the only dataset recordings bundled are the seven core-word examples above. View the official ASL-Citizen project and download terms. Dataset entries are not automatically supported camera words.
Find a sign by topic, watch an available example, or open a lesson. Recognition badges show which words have an enabled preview model. New category models did not pass evaluation. Other words can be practiced with personal recordings.
Category models load only when selected.
Loading vocabulary…
Emergency and safety entries are learning vocabulary. Recognition can miss or confuse signs, and this website does not contact emergency services.
External lessons open their publisher’s website. ASL signs and phrase grammar vary by context; English phrases are not automatically a single sign. Typed text buttons do not count as camera recognition.
Personalized machine learning
Recognize letters, then personalize your signing.
Seven letters come pretrained from ASL image examples. Start the camera to try them, or capture your own examples to personalize recognition. Teach the 5 gesture before using it. Uncertain poses are left blank instead of guessed. Camera frames stay on your device.
21 landmarksOn-device trainingNo images stored
AI training setLoading pretrained letters…
Start the camera above, hold the selected gesture, then capture samples.
Optional: add your own word examples
Record at least three examples for a new word. Removing personal examples keeps the pretrained vocabulary.
Word recording and review options
Words in motion
Sign once, pause, then review the match.
The camera suggests ten words: HELLO · THANK YOU · PLEASE · YES · NO · FINISH · DRINK · WATER · FOOD · HUNGRY. The last four have lower recorded accuracy. Review every suggestion.
Other vocabulary & recording options ExperimentalRecorded word accuracy
Loading recorded test results…
Correct test clips divided by all clips in the listed test. This is not confidence in the current camera match. Personal examples have no independently measured accuracy.
Daily conversation · five-word preview
NAME · HOME · HAVE · WANT · LOVE
Experimental: 51/66 held-out word clips matched; 3/48 unsupported signs were falsely accepted. This preview has limited accuracy. Review every suggestion. Results are not live-camera accuracy.
Video stays on your device. Recorded coordinates are processed in memory and not stored. With Automatic words enabled, sign once and pause to send a captured sequence automatically. Review each match before adding it; optional automatic adding uses the setting below. Recording manually is still available.
Loading five-word preview…
On unless you switch it off. These four come from the extra-word model: on its recorded test clips DRINK matched 7 of 8, WATER 5 of 8, FOOD 5 of 5 and HUNGRY 4 of 5, and 11 of 16 unsupported signs were falsely accepted. Its other words (BOOK, LIKE, GO) are not suggested. Untick to use the first six words only. Review every suggestion. Recognition checks
Loading DRINK, WATER, FOOD and HUNGRY…
Loading pretrained words…
Loading pretrained vocabulary…
Pretrained vocabulary will appear here.
Each recording has a 1.5-second countdown, then captures 2.5 seconds of movement. Slow cameras get up to 2 extra seconds to collect 12 frames.
Built for inclusion
Communication should feel effortless.
HandSpeak Web is a capstone prototype that uses computer vision to help bridge everyday communication between deaf, hard-of-hearing, and hearing individuals. It recognizes selected static letters and offers an experimental pretrained vocabulary for moving signs. Face landmarks are displayed for visual feedback; they are not used to interpret identity, emotions, or ASL grammar.
7Tested static letters
On-deviceCamera recognition
0frames stored
Help, privacy, and recognition limits
Letters: hold a supported letter steady. Words: select Words in motion, sign once, and pause. Watch the examples first; use Record a word if automatic capture misses the movement.
Camera frames and personal gesture examples stay on your device. The optional five-word preview sends recorded hand-landmark coordinates to the server for in-memory recognition; it does not upload video or save those coordinates. Saving sends your message and letter-detection statistics to this server. Guest history is linked to a browser cookie. Clearing cookies removes access to guest history. Existing account sessions retain access to their saved data. You can export or delete your own saved messages.
Feedback is stored on this server. Public recognition, learning, and saving do not require sign-in.
This is an experimental ASL learning and recognition application, not a complete language translator. Similar and unsupported signs can be misidentified. Extra-word recognition is a preview. Always review suggestions before adding or speaking them.