Talk to Me, Goose! is an AI-assisted communication app built by — and for — people
living with ALS. As the UX lead on a 5-person team, I conducted accessibility research, identified usability gaps across platforms, and delivered a high-fidelity prototype designed for clarity, control, and personality.
Talk to Me, Goose! was created by our client, David Betts, after his own ALS diagnosis — a way to preserve his voice and self-expression as the disease took his speech. Since launch, the app has earned a 2026 Zero Project Award and local news coverage, and now serves a community of more than 30,000 people living with ALS in the US.
David asked our team to evaluate and improve the app's usability — to make it easier and faster to use. But the deeper we looked, the clearer it became that the friction wasn't really about speed or missing features — it was about whether people felt in control of what the AI said on their behalf. That wasn't a usability problem; it was a problem of agency — and it reframed everything we designed next.
Rather than design once and test at the end, we ran three iterative rounds — discovery, mid-fi iteration, and hi-fi validation — each feeding directly into the next. And we designed with the community, not just for it: testing moved from simulated users to people living with ALS and expert speech-language pathologists.
A multi-method strategy kept the redesign grounded in evidence and lived experience.
The reframe came directly from watching people use the app. Across think-aloud sessions and a heuristic review, the same pattern surfaced: people hesitated, second-guessed, and avoided features they couldn't predict. In assistive communication, "close enough" changes meaning — users were settling, adjusting to the app instead of expressing themselves through it. The breakdowns were concrete: messages that didn't fully capture what they meant, choices they couldn't undo, and feedback that didn't confirm their actions.
Research GoalsMap the usability and accessibility gaps in the existing product.
Understand ALS users, caregivers, and family members — and how their communication needs differ.
Understand what makes someone feel in control — not just whether the app works.
Two Merlin entry points, lookalike phrase tools, and no confirmation that an action had worked.
We started with secondary research — stakeholder mapping, competitor analysis, heuristic evaluation, and a literature review of 15 AAC papers — to understand the problem space before testing with users. The literature review was clear that AAC should support a person's expression, not speak for them, which made Merlin's binary accept/dismiss the first thing worth investigating.
We used simulated users first — non-impaired participants following speech simulation guidelines — because recruiting ALS participants required more time and ethical preparation. Participants completed everyday communication tasks: composing a message with Merlin, navigating between features, and adjusting voice settings.
SynthesisWe used affinity diagramming to synthesize findings from think-aloud testing and secondary research — clustering recurring behaviors, pain points, and emotional reactions to surface patterns around usability, agency, and communication breakdown.
Raw observations
Grouped into themes
Users type short or partial input expecting Merlin to understand context. But it over-interprets — "help me up plz" becomes "I'm having trouble getting out of my chair," assuming a physical context the user never mentioned.
In Story Builder, a prominent "Start Over" button left users unsure what would reset — the whole session or just the last step. Uncertainty about recovery made users cautious, avoiding features rather than exploring them.
The main page offers one suggestion: accept or dismiss. But users wanted to swap tone, adjust intensity, or edit a phrase without regenerating everything. Story Builder gave even less flexibility — fixed options with no room to deviate.
Quick Text and Common Phrases look identical but behave differently. Merlin appears in two places and generates output in subtly different ways. Users spend time decoding the interface instead of communicating.
This is not just a usability problem.
It's an agency problem.
Users aren't just trying to complete tasks. They need to express themselves in real time — with control over what they say and confidence in how they say it.
Before optimizing for speed or expression,
users need to feel in control.
Based on our research, we explored three strategic directions and chose Safe & Confident Interaction. Testing showed hesitation and uncertainty were the real barriers to communication, so improving clarity and confidence first delivers immediate usability gains and lays the foundation that speed and expression depend on.
Lightweight Expression
Deeper control over wording and tone — flexible editing to restore a sense of authorship.
—Held for later: high impact, but more design and technical effort — a longer-term evolution.
Safe & Confident Interaction
Make actions predictable, visible, and reversible — directly answering the hesitation and uncertainty users showed in testing.
Chosen DirectionFaster than Speech
Shortcuts, reuse, and conversational momentum to speed up communication.
—Held for later: valuable, but an optimization layer that needs a stable foundation first.
We translated research into a testable prototype using an AI-assisted workflow — writing a Feature List, PRD, and Prototype Brief, then feeding them into Google AI Studio to build and iterate. The interface organized into two clear areas: a speech panel on the left and a customizable support dashboard on the right. This round also brought our first sessions with people living with ALS.
Mid-fidelity prototype: split-screen layout with speech panel and support dashboard.
The prototype kept a split-screen format, reorganized around a speech panel on the left and a support dashboard on the right.
We tested with 12 participants — two with diagnosed ALS and 10 using simulated testing protocols — reusing the same family scenarios from Round 1 for a direct before-and-after comparison. The goal was to evaluate whether the redesign improved confidence, clarity, and recoverability, and to surface remaining barriers to trust and control. Sessions were recorded and supplemented with moderator notes and direct quotes.
Key InsightsAcross participants, the redesign was consistently perceived as an improvement. Users still edited rather than accepted, though — AI outputs sometimes misrepresented tone or intent, especially from partial input.
The system was organized into levels, but users expected more incremental, piece-by-piece control. They struggled to predict what each level would actually do to their input.
Users only engaged with AI when they could easily recover from mistakes. Undo, Clear All, and Quick Text made interactions feel safe. Advanced modes were rejected when they lagged or felt hard to undo.
Users wanted more control over AI outputs but felt overwhelmed by the number of components on screen. They could identify issues but weren't sure which controls would actually help.
Testing confirmed the direction but surfaced three issues: confusion around AI assistance, lack of onboarding, and interface overload. We made four focused changes.
Replaced ambiguous levels with Manual, Construction, and Expansion. Added clear labels and subtitles. Introduced an autocorrect toggle with visible on/off states.
Added a step-by-step walkthrough of core interactions. Onboarding is re-enterable so users can revisit it at any point.
Simplified the layout by showing one section at a time. Added collapsible sections, icon-based navigation, and an adjustable number of AI suggestions.
Positioned Undo alongside primary controls. Clarified whether suggestions replace or append to existing text.
The final hi-fi prototype brings speech composition, AI support, reusable communication tools, and guided onboarding into one cohesive, tablet-first experience — every decision tracing back to the agency problem we set out to solve.
Tackling a technical challenge in the AI's behavior.
Expansion Mode had a problem: when a user's input was already nearly a full sentence, the AI responded to it instead of helping finish it. We fixed it so near-complete input returns variations of the user's own words, shaped by the chosen context or mood. The message stays theirs, not something the AI generated for them.
The design system was built around four personality pillars — Clear, Predictable, Controllable, and Reliable. A teal primary palette anchors the brand, distinct feature colours differentiate each mode at a glance, and a two-typeface system (SF Pro for UI, Montserrat for display) keeps the interface readable and consistent across contexts.
For final validation, we tested the hi-fi prototype with two users living with ALS and two expert speech-language pathologists — combining qualitative feedback with structured quantitative metrics. After earlier rounds explored broadly, this phase set out to confirm the redesign was working as intended.
David asked for an easier, faster app. Research revealed that speed alone would have missed the point. For people with ALS, the goal isn't only to be understood — it's to still sound like themselves. The final design makes the system's behavior transparent and gives users clearer control over how AI is applied, and validation showed that improving clarity and control let people focus on expressing themselves, not decoding the interface.
Why This Project MattersDavid asked us to make the app easier and faster. What actually mattered was whether people felt in control, and I'd have missed that if I'd just built what he asked for.
People only trusted the AI when they could still edit or undo what it wrote. The moment it wrote something they wouldn't say, they stopped using it.
Testing with people who actually live with ALS, and the speech pathologists who help them, turned up problems I'd never have caught alone.
