What 12 AAC Users Taught Researchers About the Future of AI-Powered Speech 👩🔬
For decades, AAC users have faced a communication tax that most people never have to pay. Augmented communicators typically produce speech at 3–20 words per minute, against 100–140 WPM for typical speakers — a gap wide enough to reshape entire relationships. People report losing the ability to joke, to add nuance, to jump into a conversation before the moment has passed. Every reply is a calculation: how much do I have to say, and can I afford the physical and cognitive cost of saying it?
Prior research has documented AAC users doing hours of prep work before a single doctor's appointment or meeting, just to have answers ready in time. This is the baseline AI has to improve on — and it's a low bar in the sense that almost anything that saves keystrokes helps, but a high bar in the sense that it has to help without taking away the one thing AAC users have fought hardest to keep: their own voice.
A team from Google Research, Carnegie Mellon, and UCL tested exactly that trade-off. They built three "speech macros" — small, purpose-built prompts on top of a large language model — and put them in front of 12 adult AAC users over live video calls to see what actually helped.
The three tools
Extend Reply — type "work," and in reply to "Do you want to get pizza?" get a full, contextually appropriate sentence like "I'm already at work" or "I'm busy with work today."
Reply with Background Information — pre-load a short bio, and let the model draw on it to answer questions like "Do you like animals?" with specific, personal detail.
Turn Words into Requests — type "bathroom" or "tea," and get a polite, ready-to-speak request.
The upside was real — and participants felt it immediately
This wasn't a lukewarm response to a nice-to-have feature. It was closer to relief. Participants described the potential to type less as directly opening up social life that fatigue currently closes off — one explained that she often doesn't even start conversations because she doesn't have the energy to type enough to keep up, and that shortcut alone could change whether she engages at all. Another pointed out that the suggestions did more than save keystrokes — they took over some of the cognitive labor of guessing what response would be socially "correct," freeing up attention that would otherwise go toward simultaneously composing a sentence and operating the device itself.
The "Turn Words into Requests" macro was the standout: for everyday needs — TV, sleep, a drink, tea and biscuits — it consistently produced usable, appropriate requests, and it was rated the most useful of the three tools by a clear margin. That's a meaningful result on its own: a huge share of AAC communication is functional and repetitive, and this is exactly the category where AI has room to eliminate real friction with very little downside.
The background-information macro showed something even more striking: genuine world knowledge doing useful work. Told that a participant was from Sri Lanka, the model inferred she might enjoy the beach — an unprompted, contextually sound guess that impressed her precisely because it felt like it understood something about her rather than just echoing her words back. Another participant watched the model correctly identify that his favorite soccer team was based in Guadalajara and build that detail into a full sentence — a level of inference no static word-prediction system could offer.
And crucially, participants weren't just impressed as observers — they immediately started imagining real use cases beyond the study: prepping for medical appointments so nothing gets lost under stress, ordering at a regular coffee shop, answering the "how are you?" they get asked dozens of times a day without having to type the same reply from scratch every time. That's the clearest signal that this isn't a novelty; it's solving problems people already have.
What still needs work — and why it's solvable
None of this means the tool was finished. The model sometimes filled in specific, unrequested details, guessed personal facts incorrectly, and occasionally missed high-stakes requests like medical suction or mobility transfers, defaulting to generic or unrelated phrases instead. A built-in safety filter also blocked a legitimate request in one case, redirecting a participant away from something he had every right to ask for himself.
The through-line in all of this feedback, though, isn't "this doesn't work" — it's "this needs to work for me specifically." Participants wanted the ability to feed in medical details, relationship context, and personal shorthand (many already had their own systems, like typing "1y" for yes or "WON1" for wonderful) so the model could learn their world instead of guessing at it generically. That's a solvable design problem, not a ceiling on the technology — and it points toward AI-enabled AAC that's tuned to the individual rather than a one-size-fits-all predictor.
Participants also had clear, reasonable conditions for adoption: real encryption, real control over when the system is drawing on live context, and the ability to edit or override any suggestion before it's spoken. None of these are objections to the concept — they're the product requirements for building it responsibly.
The bigger picture
What this study captures is a group of expert communicators — people who have spent years optimizing around a genuinely difficult input problem — looking at a new tool and seeing a path to something better than what they've had. Not a replacement for their own voice, but a way to reclaim time, energy, and spontaneity that the current generation of AAC devices structurally can't offer. The researchers' own framing is telling: participants didn't just want less typing, they wanted the right less typing — customizable, personal, and theirs. That's a far more exciting design target than simple word prediction, and this study is early evidence that large language models can actually hit it.