The Sentence Engine
A child who can build sentences needs words — fast.
By Tier 3, a child is constructing 3–4 word sentences. They know what they want to say — “I don’t want the blue one” — but finding each word across a large vocabulary is slow and frustrating.
Proloquo2Go solves this with 200+ folders. But folders are a taxonomy — abstract, hierarchical, hard for young children to navigate. Meadow needs a different answer.
Meadow’s Tier 3 paradigm
Instead of matching Proloquo2Go folder-for-folder, Meadow delivers rich vocabulary through two innovations that play to its strengths.
The Sentence Engine
Grammar-aware predictions appear inside the speech bubble as the child builds a sentence. The app understands what kind of word comes next — a verb after a pronoun, a noun after a verb — and surfaces the most likely options. Fewer taps. Faster sentences.
- ✓ Rule-based, fully offline
- ✓ Ranked by scene context + usage
- ✓ Replaces rigid SCS templates
World Map Expansion
Every vocabulary domain maps to a place on the world map. Animals live at the zoo. Body parts are at the doctor’s office. Clothes are in the bedroom. The map replaces folder hierarchies with spatial memory — children remember where words are.
- ✓ Max 2 taps to any word
- ✓ 15–20 illustrated scene locations
- ✓ 700–1,500+ words V1
The speech bubble becomes a sentence builder
At Tier 2–3, the speech bubble does double duty: it shows the words the child has composed and predicts what comes next. One unified surface for sentence construction — works in both scene view and grid view.
Always optional
Predictions never replace manual word selection. The child can ignore predictions entirely and tap any word on the compass edges, scene, or grid — it gets added to the bubble just the same. Predictions are a shortcut, not a requirement.
How it feels to build a sentence
Four scenarios showing the engine responding to different sentence starters and contexts.
“I don’t want…”
Child taps I → don’t → want. Engine predicts nouns because Negation+Verb+Object is the pattern. Kitchen foods surface first.
“big dog!”
Child taps big → engine predicts nouns. In the zoo scene, animals surface first. “Dog” ranks high because this child uses it frequently.
“no!”
Child taps no. Engine predicts verbs (the child is protesting an action). Bath-context verbs rank first.
“where mommy?”
Child taps where. Engine predicts nouns and pronouns — the child is asking about a person or thing. Family members rank high.
Rule-based. Offline. No machine learning.
The engine uses the Fitzgerald Key color of each word as its grammar role. A small transition table maps what’s in the bubble to what category of word probably comes next. The entire system runs on-device with zero latency.
| After… | Predict next | Example |
|---|---|---|
| Empty bubble | Pronoun Verb Noun | “I…” “want…” “banana” |
| Pronoun | Verb Descriptor | “I want” “mommy happy” |
| Verb | Noun Descriptor Function | “want banana” “want more” “want to” |
| Noun | Descriptor Social Function | “banana yummy” “cookie please” “ball and” |
| Negation | Verb Noun | “don’t want” “no banana” |
| Question | Noun Verb Pronoun | “what that” “where daddy” |
Grounded in Brown’s Stages II–IV
These transitions map to the documented sentence patterns for children ages 24–48 months (MLU 2.0–4.0): Agent+Action, Action+Object, Negation+Verb+Object, and early question forms. The transition table will be validated by SLP review before implementation.
Context-aware, personalized, diverse
Once the grammar engine selects a category (e.g. “show nouns”), the ranking system decides which nouns appear. Four layers, highest priority first.
Scene context
Words from the current scene location rank highest. In the kitchen, kitchen objects appear first.
Usage frequency
Words this child taps often rank next. If “ball” is their most-used noun, it appears even in the kitchen.
Recency
Words used in the last few minutes get a boost. Supports ongoing conversation topics.
Pragmatic diversity
At least 2 predictions must serve different communication functions — not all requesting. A child who said “I want” also sees commenting and social options.
On-device, privacy-preserving
All ranking data comes from the existing SwiftData event log. No data leaves the device. No cloud processing. No usage analytics transmitted anywhere.
One system, two intensity levels
The sentence engine scales with the child. Tier 2 gets a gentle version. Tier 3 gets the full experience.
Simplified predictions
- ✓ 3–4 predictions shown
- ✓ First word only — no multi-position chaining
- ✓ Scene nouns only — no cross-scene vocabulary
- ✓ Larger tiles (60pt) for developing motor skills
- ✓ Tapping a prediction speaks the full 2-word phrase and clears
Full predictions
- ✓ 6–8 predictions shown
- ✓ Multi-position — predictions update after each word
- ✓ Full vocabulary — cross-scene, usage-ranked
- ✓ Tighter tiles (50–55pt) for better motor skills
- ✓ Max 5-word composition with scrollable bubble
Every vocabulary domain is a place
Instead of folder hierarchies, Meadow organizes vocabulary as locations on a world map. Children remember where things are. The zoo has animals. The doctor’s office has body parts. The bedroom has clothes. Max 2 taps to any word — flat, never nested.
Kitchen
Bathroom
Bedroom
Living Room
Zoo
Doctor
Park
Grocery Store
Grandma’s House
Car Ride
Art Room
Restaurant
Beach
Garden
The child’s dog lives at the zoo
SLPs and parents can add personal vocabulary — pet names, sibling names, favorite shows — directly into the scene where they belong. No separate “personal board.” The child’s dog appears in the zoo scene alongside “dog,” with a photo of their actual pet.
Parent takes a photo
Camera roll or camera. The real dog, not a generic illustration.
Apple TTS speaks it
On-device text-to-speech. Instant, offline, no pipeline needed.
Appears in the right scene
SLP picks the scene. Custom words have a subtle photo-frame border.
Taylor’s popsicle
In Taylor’s testimonial, specificity was the breakthrough — not “frozen treat” but the exact color popsicle she wanted. Custom words enable this kind of specificity without requiring Meadow to pre-build every possibility.
Clear boundaries
The sentence engine is a communication tool, not a teaching tool.
Not AI / ML
Simple rule-based grammar transitions. No neural nets, no cloud calls, no black boxes.
Not a lesson
Predictions help children communicate, not learn grammar. No teaching, no correction, no assessment.
Not mandatory
Every word is still directly tappable on compass edges, scenes, and grids. Predictions are a shortcut, never a gate.
What it is: the fastest path from thought to speech
A child who knows what they want to say should never be slowed down by the interface. The sentence engine anticipates their next word, surfaces it where they’re already looking, and gets out of the way. The child builds the sentence. The engine just reduces the taps.
The vocabulary platform (V2)
The world map architecture sets up a platform play that no competitor has built.
SLP-authored vocabulary packs
SLPs create new scene locations with vocabulary, illustrations, and audio. Packs are shareable through a curated library. Installing a pack adds a new location to the child’s world map. Every SLP who creates a pack makes Meadow more useful for every child.
Competitive gap validated
No AAC app — including Proloquo2Go, TouchChat, TD Snap, or CoughDrop — has an SLP-authored vocabulary pack marketplace. SLPs currently sell custom materials on Teachers Pay Teachers as a workaround. Meadow would be the first app where the scene-based paradigm makes vocabulary packs genuinely rich (not just a grid layout, but an illustrated world with vocabulary, context, and therapy targets).