Turn

A question. Your words: pick a place, and your own saved replies are ready.
Consent comes first: you allow it, then your partner agrees.
Tap a phrase, or type: speaking is always free.
Your words, your way: edit every phrase, with Ren beside your grid.
Turn Listen: one $24.99 payment through RevenueCat, with Restore Purchases in Settings.
A face for your voice: choose a companion, and show your partner what you said.
Screenshot, 1179 × 2556, no device frame: "How was physio?" with the saved reply "It was hard"
App icon, 1024 × 1024, uncropped

Entering: the Next Gen Award. Turn is a student team's entry, submitted with a public repository and a demo video instead of a store listing.

What it does

Turn is an augmentative and alternative communication (AAC) app for iPhone. It's for adults who can read and tap but can't rely on their speech, such as people living with ALS or recovering from a stroke.

Speak. Tap a saved phrase or type, and Turn speaks it in your Personal Voice or a system voice. The bank starts with about 150 editable phrases in categories, and every typed reply joins it.
Listen. In Listen mode, Turn transcribes the conversation partner on the iPhone. When the partner finishes a line, a row above the grid offers your own saved phrases that answer it: one big button when Turn is confident, up to six otherwise, and no change when nothing fits. For a yes-or-no question, Yes, No, and Not sure come first.
Nothing speaks until you tap. A reply keeps its slot until a new one wins by a clear margin, and the grid never reorders itself.
The "aha". A partner asks "How was physio?", and "It was hard" is waiting in the row, though the two share no content word.

Why it matters

Aided communication runs at about 8 to 10 words a minute, while speech runs at 125 to 185. By the time a reply is typed, the conversation has moved on. Existing AAC apps sell symbol grids or typing. Newer apps that listen generate replies, and AAC users testing AI suggestions worried about the system "suggesting the wrong thing and making the participants look bad". Turn never writes a word: a hosted decision model chooses which of the person's own phrases answer each line, and whether any do.

The evaluation

On September 23, 2026, we evaluated 80 partner lines with the hosted decision model pinned to version 1.13.0. On the 64 lines with an acceptable saved reply besides Yes, No, and Not sure:

Ranker	Top 1	Top 6
chance	5%	26%
keyword	16%	23%
embeddings	22%	45%
hosted decision model	67%	75%

Its top 6 leads embeddings by 29.7 points (95% paired interval, 17.2 to 42.2). Claude subagents wrote the lines, the starter bank, and the labels at our direction, and no clinician has reviewed Turn yet. The repository's eval/results.md has the intervals and caveats.

How it makes money

Speaking is always free: the grid, typing, saved phrases, and Personal Voice. Turn Listen, a one-time $24.99 purchase through RevenueCat that grants the listen entitlement, keeps Listen mode on after 20 free partner lines. Listen mode is the part that costs us money on every line, and a one-time price means nobody loses their voice when a payment lapses.

The RevenueCat Paywall opens on the 21st partner line. Settings holds Unlock Listen mode and Restore Purchases.
Our relay checks the listen entitlement through RevenueCat's REST API v2 before it calls the model.
The purchase runs through RevenueCat's Test Store, which the organizers accept for Next Gen: "Test Store is enough for the Next Gen category" (forum). Test Store purchases are sandbox data, so we report no revenue.

How judges reach the paid part

Download Turn.app.zip from the latest release, then install it on an iOS 27 Simulator with xcrun simctl install booted Turn.app.
Tap Home at the top of the grid, choose Clinic, then tap Listen, Allow, and They agreed.
Type "How was physio?" as the partner's line, and tap "It was hard" to hear it.
Open Settings, tap Unlock Listen mode, and complete the Test Store purchase. It's free, so no promo code is needed.

Bundle ID: com.m1ku.turn.

How we built it

App: Expo SDK 57, React Native 0.86, TypeScript, Expo Router, and expo-sqlite, with our own Swift Expo modules for on-device transcription with Apple's SpeechAnalyzer and for Personal Voice.
Purchases: react-native-purchases and react-native-purchases-ui, with a RevenueCat Paywall configured remotely.
Relay: a Cloudflare Worker with SQLite-backed Durable Objects. It counts free lines, checks the entitlement, rate-limits, and asks the model to score a 40-phrase shortlist.
Retrieval before judgment: the phone picks 40 candidates by keyword ranking, the place, and your most-used replies, so the whole phrase bank never leaves the phone.
Offline: without a network, or when the model is slow or down, the phone ranks phrases itself and says Listen mode is degraded.

Privacy and consent

Before any phrase leaves the phone, you give permission. Then the partner sees a consent card each time listening starts.
A light shows while Turn listens, and a switch keeps a partner under 18 from being heard.
There are no accounts and no audio is recorded. Names become tags before a line leaves the phone. The privacy notice is also in the app.

AI tools we used

We built Turn with Claude Code and OpenAI Codex as coding agents, and Claude subagents drafted the evaluation lines, starter phrases, and labels. We set the direction, reviewed the work, and tested it on an iPhone and in the Simulator.