Is Oso Voice's Word Navigation Going the Wrong Way?

Oso Voice research — 2026-07-28. Four parallel high-effort research threads on TouchChat WordPower, PRC’s Minspeak/Unity, and the wider AAC evidence base — checked against Oso Voice’s current fixed-paged-grid design, to answer a specific worry: that words should carry some correlation to each other, and right now they don’t.

Copied into the repo 2026-08-09 from the session artifact “Navigation Deep Dive — WordPower, Minspeak & the Evidence Base” (2026-07-28/29), per the 2026-07-20 cell-color-formatting precedent. This document informed the approved spec docs/superpowers/specs/2026-07-28-hub-and-spoke-navigation-design.md. Confidence labels: [peer-reviewed] [vendor] [secondary] [not found].

Headline

The architecture — flat paging, one word one fixed position, no folder trees — is not the problem. It’s unusually well aligned with the evidence. The real gap is exactly what was named: nothing today makes what’s available depend on what was just picked. That’s fixable without touching the parts that already work, and the field has a well-precedented, low-risk pattern for it — it’s just not the pattern (Minspeak icon-sequencing) that “look at WordPower” first suggests.

01 — What the evidence says you already got right

Before changing anything, it’s worth being precise about which parts of the current design the research actually backs — so they don’t get disturbed by mistake while fixing the real gap.

Flat, shallow paging [peer-reviewed] — Light & Drager (2007) found preschoolers had more trouble picking the right page out of 4 unseen options than the right symbol out of 12–30 visible ones. Petroi et al. (2014) found accuracy and speed both drop as navigation depth increases — but in adults with Broca’s aphasia (PubMed 24575783), not children, so it carries to this population by inference only; Light & Drager (2007) is the child-specific half of this finding. Adding folder levels would move against this, not with it.
Positions that never move [peer-reviewed] — Thistle et al. (2018) is the strongest single study in this space: children with stable symbol locations got faster over 5 sessions; children with reshuffled locations didn’t improve at all. Oso Voice’s one-word-one-permanent-position rule passes this test directly.
The core-words toolbar [vendor-precedented] — Already ships, always visible regardless of which grid page is showing. This is structurally the same mitigation WordPower’s own design goal describes (“core words remain consistent… a natural bridge between core and fringe”) — just implemented as a strip instead of embedded duplicate buttons.
Grid-size change warning [peer-reviewed-relevant] — Measuring WordPower’s own published layouts (WP42 Basic vs. WP60) shows most core words do move position when the grid size changes — its “motor planning built in” claim doesn’t survive a size change either. The warning already shown on grid-size change is addressing a real, evidence-matched risk.

02 — The actual gap: no correlation mechanic exists yet

Right now, picking a word never changes what’s offered next. Every page’s contents are static, hand-curated, and indifferent to the message being built. That’s the concrete shape of “we’re missing how words carry correlation” — and every AAC system researched does some version of fixing it. They differ enormously in mechanism and risk, which is the part worth getting right before building anything.

Tier	Mechanism	Precedent	Risk for this app
1	Dynamic next-word strip — core grid stays fixed; one row repopulates with contextually likely next words after a selection	Unity 1-Hit “activity row”; WordPower’s “logical next words” prediction cells	Low
1	Verb → category auto-jump — picking a core verb like “eat” or “want” navigates straight to the relevant topic page	WordPower’s embedded intelligent links (eat→FOOD, drink→DRINKS, play→TOYS); Grid 3’s “Intelligent Jumps”	Low–medium
2	Positionally-consistent branch pages — if a tap opens a sub-page, the verb/adjective/noun slots sit in identical coordinates no matter which word opened it	Unity Sequenced’s layer-2 grid (verb always same cell, adjective always same cell, regardless of parent icon)	Medium
3	In-place grammar morphing — the same button relabels itself based on sentence context (“go” → “going”)	Proloquo2Go Dynamic Morphology, Grid 3 Smart Grammar	High
3	Full icon-sequencing (semantic compaction) — one icon means different words depending on the icon sequence pressed before it	PRC Minspeak / Unity Sequenced	High — avoid
—	Reordering the main grid by recency or frequency	Not found shipping in any mainstream AAC system	Anti-pattern

Why Tier 3 is the wrong instinct here, even though it’s the most famous version of “dynamic”

Minspeak/Unity’s icon-sequencing is the historically dominant answer to “meaning depends on what came before” — a small icon set where sequences disambiguate to specific words. It’s the mechanism that most closely matches the literal words “words carry correlation.” But it was tested directly on the population closest to Oso Voice’s:

Typically developing 4–5 year-olds scored 14–27% accuracy with iconic/sequenced encoding, against 62–80% for taxonomic grids, schematic grids, and visual scenes — and children “generalized to novel vocabulary items except in the iconic encoding condition.” — Light, Drager, McCarthy et al. (2004), summarized in Light & Drager (2007), Penn State (full text)

A follow-up (Drager & Light, 2010) tested whether greying-out invalid next icons (“icon prediction”) rescues this — it found no significant learning benefit. Autistic and speech-delayed children are a harder population than the typically-developing kids in these studies, not an easier one. Van Tatenhove’s own parent-training guide for Minspeak requires teaching parts of speech as a prerequisite — metalinguistic knowledge SLP literature says most children don’t have until 5–7 years of fluent talking. This is a real, well-evidenced reason to route around Tier 3, not just a hunch.

03 — What “how words relate” looks like in WordPower specifically

WordPower is the more directly relevant precedent for the correlation mechanic, since — unlike Minspeak — it’s a plain word-based system philosophically close to Oso Voice already (Nancy Inman built it explicitly because her first user “did not want to learn the complex language coding system” of icon sequencing).

The organizing axis switches with depth

WordPower’s home page is ordered by grammar — Fitzgerald left-to-right (pronouns → verbs → descriptors), color-coded by part of speech. Category pages one level in switch to a semantic axis, and color is explicitly repurposed there for visual discrimination rather than grammar. Nouns are evicted from the home page entirely, reachable only through a GROUPS button.

Core is embedded, not duplicated wholesale

The actual mechanism for “the same motor pattern works whether you’re on the home page or three taps into a category” isn’t a fixed spine repeated everywhere — measuring PRC’s own published WordPower60 screenshots shows it’s a curated ~16-cell verb block that keeps its coordinates on some pages (confirmed on the ACTIONS page) and is completely absent on others (the EXTRA WORDS page has zero overlap). The vendor’s own language hedges this precisely: [vendor] “motor planning is built into much of the file,” not “every page.”

The actual correlation mechanism: verb-triggered jumps

This is the part worth borrowing. Specific core verbs are hard-wired to jump straight into their topic page — eat→FOOD, drink→DRINKS, play→TOYS/SPORTS, listen→MUSIC, watch→TV/MOVIES, wear→CLOTHES, go→PLACES. Smartbox’s Grid 3 implementation calls this “Intelligent Jumps” and describes it plainly: if you’re building “I want to drink…”, you jump straight to the drinks grid without an extra tap. This is structurally simple, reversible, well-precedented across two independent vendors, and is the most direct match to “navigation dynamic on previously selected words.”

Color scheme: WordPower is not Modified-Fitzgerald-Key compliant — it uses pink for adjectives where MFK uses blue, and blue for function words where MFK uses grey. Oso Voice’s existing MFK-based ADR would clash if this got borrowed alongside the navigation ideas.
The efficacy claim itself: no peer-reviewed study evaluates WordPower’s navigation design at all. The one strong outcome study (an RCT, MLU +43% vs. +6%) used WordPower only as the delivery vehicle for a therapist-led intervention with deaf/hard-of-hearing children — not evidence that its organization beats alternatives.

04 — How sure is any of this, really?

Worth being honest about tiers of confidence before this turns into product decisions — the field asserts more certainty than it has evidence for.

[peer-reviewed] Fixed symbol location speeds learning over repeated sessions (Thistle et al. 2018); page-navigation errors exceed within-page search errors (Light & Drager 2007); iconic/sequenced encoding underperforms 3 alternatives for TD 4–5yo (Light et al. 2004).
[vendor claim] “Communication rate up to 6× with Minspeak”; “90% of communication uses semantic compaction” — both trace to an unpublished dissertation and ASHA conference talks, not journal articles, co-authored by PRC’s own co-founder.
[secondary] Clinical blog commentary, SLP training guides, App Store reviews (used here only to corroborate a directly-measured vendor screenshot, never standalone).
[not found] Any direct study comparing hub-and-spoke to sequential paging; any independent evaluation of WordPower’s navigation model; any study isolating “navigation design” as the active ingredient in LAMP’s outcome studies.

The practical read: treat any claim that “research proves WordPower/Minspeak’s navigation is better” — from a competitor, a reviewer, or future-us — as unsupported. It doesn’t exist yet. What does exist points at the narrower, more useful conclusion above: keep the shallow flat structure, keep fixed positions, and close the correlation gap with a verb→category jump plus a contextual next-word strip, not a structural rebuild.

05 — Suggested next step

This is a research synthesis, not a spec — no code has been touched. The natural next step is superpowers:brainstorming to scope which of the Tier 1 mechanisms to build first (the verb→category jump is the smaller, more self-contained change; the next-word strip touches more of the message-building flow) before writing a design spec the way the fixed-paged-grid work was scoped in 2026-07-21.

(2026-08-09 note: what actually shipped first from this research was the hub-and-spoke navigation redesign — the chrome-cost half of the findings. The Tier 1 correlation mechanics above remain unbuilt and are still the recommended next mechanic; see also Adam’s 2026-08-09 mixed-Home observation, which aligns with §03’s “organizing axis switches with depth” finding.)

Primary sources referenced

Light, J. & Drager, K. (2007). AAC: State of the Science and Future Directions. Penn State, open access
Thistle, Holmes, Horn & Reum (2018). Consistent symbol location affects motor learning in preschoolers. AJSLP 27(3). PubMed 29860450
Drager & Light (2010). Icon prediction and iconic encoding learnability. AAC 26(1). PubMed 20196700
Petroi, Koul & Corwin (2014). Navigation depth effects on accuracy/latency. AAC 30(1) — participants were adults with Broca’s aphasia, not children. PubMed 24575783 · ResearchGate
WordPower Vocabulary Quick Reference, PRC-Saltillo (2025). touchchatapp.com
WordPower Manual for TouchChat/NovaChat/ChatFusion, N. Inman (2023). touchchatapp.com
Features of WordPower — Smartbox Hub. hub.thinksmartbox.com
Minspeak History — PRC. minspeak.com
Unity Vocabulary Quick Reference Guide, Liberator/PRC (2025). liberator.co.uk
AssistiveWare — Dynamic Morphology gating rationale. gatewaytolanguageandlearning.com
Carnett, Devine et al. (2025). Systematic review of core-vocabulary AAC interventions (WordPower absent). Review Journal of Autism and Developmental Disorders 12(2). full text
PODD deliberate vocabulary duplication rationale — Jabbla / CPEC. jabbla.com