Ugandan Sign Language
Speech becomes sign, while the speaker is still talking.
VoxSign listens to spoken Luganda or English, transcribes it with a speech model fine tuned on Luganda, matches the words against a reviewed Ugandan Sign Language vocabulary, and plays the signs back on a 3D avatar in time with the audio. Captions run alongside, for the people in the room who read rather than sign.
Free to start. No card required. Built at Makerere University in Kampala.
In partnership with
- MIICMakerere Innovation and Incubation Centre
- Where VoxSign is built, at the College of Computing and Information Sciences.
- UNADUganda National Association of the Deaf
- The Deaf community organisation VoxSign works with on Ugandan Sign Language.
How it works
Four steps, and the timing is kept throughout.
The point of a translation is that it lands at the same moment as the speech it came from. Every stage below carries timing forward rather than throwing it away.
You speak, or you upload
Talk into the app for a live translation, or upload a recording of a lecture, a meeting or a clinic visit for a learning session you can replay.
The audio is transcribed
A Whisper speech model fine tuned on Luganda transcribes short segments of the audio at a time, so it follows how people here actually speak rather than only textbook English.
Words are matched to real signs
The transcript is matched against a reviewed Ugandan Sign Language vocabulary. Each match becomes a cue with a start and an end time, tied to the moment the words were said.
The avatar signs it back
A 3D avatar performs those cues in order, synchronised to the original audio or video, with captions running underneath so sign and text arrive together.
What it does
Built around the moments where hearing is assumed.
A classroom, a clinic or a staff meeting should not stop being understandable the moment somebody starts talking.
Live translation
Press to talk and watch the avatar sign what you said. The transcript appears with it, so you can check the words were heard correctly before you rely on them.
Learning sessions
Upload a video and watch it again with the avatar signing alongside, the transcript scrolling in step, and the captions on. A recorded lecture stops being lost to anyone who cannot hear it.
Captions on everything
Sign and text serve different people in the same room. Captions can be turned on or off and sized independently, and they scale with your text size preference.
Settings that follow you
Text size, reduced motion, high contrast and colour blind support are saved to your account rather than to one browser, so the app arrives set up the way you need it on any device you sign in from.
Accessibility
Accessibility is the product, not a settings page.
An accessibility tool that is awkward to use has not solved anything. These are in the app now, and they are testable.
- Text that scales
- Three text sizes, applied at the root, so every measurement in the interface grows with them. Nothing is pinned to a pixel size that ignores the setting.
- Reduced motion by default
- Animation is off unless you ask for it. Loading states are designed to still read as loading when nothing is moving, rather than freezing mid-spin.
- High contrast mode
- Black text on white with strengthened borders throughout, for readers who need more separation than the default palette gives.
- Colour blind support
- A colour adjustment for deuteranopia and protanopia, and colour is never the only thing carrying a meaning anywhere in the app.
- Full keyboard operation
- Every control is reachable and operable by keyboard with a visible focus ring, and an easy navigation mode enlarges hit targets and that ring further.
- Screen reader semantics
- Real headings, labelled controls, and every wait announced with words rather than left to a silent spinner.
The full detail, including the standard we hold ourselves to and how to report something that does not work for you, is in our accessibility statement.
Built for
The rooms where this actually matters.
Deaf and hard of hearing people
The people this is for. Spoken language you would otherwise have to ask somebody to repeat, arriving as sign and as text at the same time.
Teachers and lecturers
Record a lesson once and hand back a version a Deaf learner can follow at their own pace, with the signing and the transcript kept in step with the video.
Clinics and public services
The short, high stakes exchanges at a counter or a consultation, where an interpreter is not in the room and writing notes back and forth is slow.
Hearing people learning USL
Say something and see it signed. Watching the avatar perform your own words is a faster way in than working through a word list.
Honest limitations
What VoxSign cannot do yet.
Anyone deciding whether to put this in front of a Deaf learner needs to know where it stops. So here it is, on the front page.
- The sign vocabulary is small, and growing slowly on purpose
- VoxSign only plays a sign that is already in its reviewed vocabulary. That vocabulary is early. Where a word has no sign yet, the avatar stays idle and the transcript still shows the words, so a gap is visibly a gap. It grows through signers and interpreters reviewing each addition, which is slower than generating them and is the reason we will not do that.
- We never invent a sign
- There is no fallback, no approximation and no nearest match anywhere in the system. A wrong sign is worse than no sign, because a wrong sign teaches the wrong thing and is believed.
- Ugandan Sign Language first, Luganda and English in
- The speech model is fine tuned for Luganda and handles English. Other Ugandan languages are not supported yet, and other sign languages are not in scope.
- Uploads are capped at 200 MB and 20 minutes
- A learning session takes one video at a time within those limits, and processing takes a few minutes for a long recording. Live translation is for short exchanges, not for transcribing a full lecture as it happens.
- Caption timing is close, not exact
- The speech model returns the words it heard but not the moment each one was said, so captions are positioned across short segments of audio rather than measured word by word. Expect them to be near the speech, not locked to it.
- This is an assistive aid, not an interpreter
- VoxSign is not a qualified sign language interpreter and must not be relied on where accuracy is legally, medically or financially consequential. Where a decision matters, use a human interpreter.
If a sign you need is missing, tell us. The vocabulary grows through people who use Ugandan Sign Language every day, so getting in touch is genuinely how it gets bigger.
Try it with your own voice, and see where it helps.
An account is free and takes an email address. If you work with Deaf learners, run a service that should be accessible, or sign Ugandan Sign Language and want a say in the vocabulary, we would rather hear from you than guess.