SIDE QUEST / HELPING PEOPLE COMMUNICATE / AAC
AAC (augmentative and alternative communication) means tools that help people who can't easily speak, often by tapping picture symbols. But symbols can force a person to talk in a squashed form, like I · GO · BATHROOM. A Voice Before Words explores adding language AI (LLMs and NLP) that works out what the person means. It then suggests full, natural sentences, like “I need to go to the bathroom.” or, when the context fits, “I went to the bathroom.” The person picks the one that actually says what they mean. The goal is for them to speak smoothly, without sounding robotic just because they use symbols.
This browser demo shows the idea, not the final AI. The planned iOS app will use language models to turn short strings of symbols into clear, natural sentences. People will be able to pick a human-sounding voice, and use their eyes instead of their hands on Apple devices that support it.
A string of symbols doesn't have to be read out word for word. Language AI that knows the context can suggest sentences that make sense, while keeping what the person means.
→ I need to go to the bathroom.Pick the ideas you mean. No need for perfect grammar.
It suggests a few clear sentences, like present, past or a request. The person picks the one that says what they really mean.
Future AI voice models should sound more natural, in a voice the person picks.
The planned iOS app is meant to add eye control for people who can't reliably use touch, on devices that support it.
The planned language layer uses AI language models to turn a short string of symbols into a few clear, correct sentences. It should handle past and present, little words like “the” and “to,” requests, feelings and what's being talked about. The final choice always stays with the person speaking.
The speech button on this page is just a quick tech demo. The real app is meant to use better speech models. Each person could pick a voice that sounds the way they want to sound. No one should be stuck with one robot voice as their identity.
The full app is planned for the iOS App Store someday. On Apple devices that support it, the app is meant to let people control it with their eyes. That way, people who can't reliably use their hands can still pick symbols and talk. The code already explores a native Apple app with eye tracking, built in SwiftUI (Apple's app toolkit). To be clear, eye control does not work in this browser demo.
An AI model can produce smooth sentences the person never meant. That's why the app should offer choices instead of quietly rewriting what someone says. The person stays the author. The software gives them more ways to say it, and they approve what gets spoken.
This site shows how it would work and where it's headed. The goal is a much more capable iOS app. It would have AI that turns symbols into natural sentences, better voices to pick from, and eye control using the device's hardware. Before anyone relies on it every day, it has to be usable by people with different disabilities, private, fast, and safe when something goes wrong.