EXPERIMENTAL AI RESEARCH
SignBridge is an in-development prototype exploring AI-assisted American Sign Language translation. It is currently focused on recognizing isolated ASL signs—not continuous conversation or interpretation.
Explore the prototype
Research project · active development
Why SignBridge?
Why SignBridge
I created Signbridge to explore accessibility within the ASL and Deaf community. I noticed many online translators supported numerous spoken languages, but almost none had a translation for American Sign Language. As AI became more advanced, I wanted to use it to create a prototype that explores how technology could make communication more accessible to everyone no matter the language they speak.
02 — HOW SIGNBRIDGE WORKS
01
Input
Type a word in English into the box.
02
The prototype evaluates the word using Ai and converts it to a useable prompt for an Ai generative video model to use.
03
Output
The final result is an Ai generated video of your filmed signer.
03 — TRY THE PROTOTYPE
This is an early live prototype. For the clearest results, type ONE isolated word into the box and press generate ASL video. Inaccuracy may happen. SignBridge is currently only applicable for one word signs.
Open SignBridge in a New Tab
DEVELOPMENT PROCESS
A living record of the work.
For many months the sign bridge was just an idea. Before I could build it, I needed to determine which tools and technology would work best. My original goal was to translate videos of ASL being signed in English using Ai. I experimented with python and MediaPipe handtracking, but encountered limitations. The model relied on a small training set and The motion capture often misinterpreted signs, as it was specific to speed and positioning. I realized this approach to a translator may be beyond what I was capable of accomplishing with the technology and data I had available.
Instead of abandoning the project, I changed my approach to a translation system that translates English words to ASL. After testing AI models and different Ai assisted coding platforms I found a combination that produced the most promising results. Sign bridge remains an experimental prototype that is making improvements to accuracy and speed.
01
The Start
SHORT SCREEN RECORDING
December 2025
During the beginning stages of testing my original idea I attempted to create my own library of signs. I used Python and Mediapipe motion tracking to capture signs. The goal was to store this landmark data to train a Python model to recognize these signs from videos.
04
SHORT SCREEN RECORDING
February 2026
After experimenting with different coding applications and changing my approach to the project, I found that Vibe Code exhibited all the features I needed while also displaying speed. I started with a prompt that included a work flow in which GPT 5.2 analyzed an English word. GPT then converted the word to a series of actions in a script form. The motion script was divided into different sections including start of the video, middle, and end. It also provided separate instructions for the signer’s dominant and non dominant hand. For example, when the user typed “I am hungry” the script directed the signer to form a curved C-handshape with the dominant hand, position it near the upper chest with the palm facing inward, and move it downward toward the stomach. It separated each parameter of the sign including the handshape, position, location, movement, and facial expression. The script was imputed into Sora Ai to produce the filmed signer.
05
Motion Instruction Testing
March 2026
I found that describing the signs as if they were actions improved the results, but the video AI still produced pauses and uncertain movements between signs. I simplified the instructions while keeping the most important parameters such as hand shape, placement, and movement. The video above shows the most successful attempt because it included the signs ME and HUNGRY in it. However, there were pauses in between signs to return to starting position that made the video unclear and the translation messy.
06
Isolated Sign Testing
SHORT SCREEN RECORDING
June 2026
After days of testing the best script for the video, I concluded that the prompt “A person accurately performs the standard ASL sign for [WORD]” worked best. For example, “A person accurately performs the standard ASL sign for HUNGRY” generates the cleanest version of a person performing the sign without pauses or hesitation. However, the results weren’t always accurate. Most times it appeared that the basic parameters of the sign were attempted, but the sign itself was not executed accurately. I also was able to figure out the combination of two signs,but later took this feature out to focus on accuracy, as not all combinations of signs gave accurate output.
07
Interface and Deployment Progress
August 2026
I tested other methods that showed potential for future development. In the video above I used a simple motion based prompt that expressed only the most important parameters without including too much detail.
“A person has their left hand open and turned upward with the palm facing up. Their right hand forms a thumbs-up and rests on top of the left palm. Keeping both hands in this position, the person raises both hands upward together in one smooth motion to accurately perform the standard ASL sign for “help.” Both hands remain clearly visible throughout the movement.” This method successfully generated the sign for HELP, but only worked for a limited number of signs and was not consistent.
08
Latest Prototype Test
August 2026
I ultimately decided to use the prompt : “A person accurately performs the standard ASL sign for [WORD]”. It produced the most consistent results across a variety of signs. When I deployed the website, I had minor complications because the generated video had less accuracy when deployed compared to in the Ai assisted coding app. I am still working to overcome this issue but the parameters for most signs are attempted correctly and basic signs such as hello, you, and I are all accurately signed. As Ai becomes more advanced, more improvements will be made to SignBridge to improve overall accuracy and speed.
05 — CURRENT LIMITATIONS
SignBridge does not understand continuous signing, regional variation, facial grammar, conversational context, or the full complexity of ASL. Outputs may be incomplete or incorrect and should never be used for medical, legal, educational, or safety-critical decisions.
07 — CONTACT
SignBridge is looking for feedback from ASL users, accessibility researchers, and thoughtful collaborators interested in more accountable communication tools.
signbridgeproject@gmail.com
Contact the project