Confrontation-naming with a three-step semantic cue ladder — category, use, properties — recorded per item as the highest cue level the user needed. Generalisation to untrained items in the same semantic field is tracked via the Layer 2 probe.
Boyle & Coelho, AJSLP, 1995. Efstratiadou, Papathanasiou, Holland, Archonti & Hilari, JSLHR, 2018 (systematic review).Evidence base
The protocols, the measurement framework, and the citations.
StrokeVoice doesn’t invent techniques. It is a practice companion built around the established, peer-reviewed approaches the SLT community already uses for aphasia, packages them as repeatable home practice, and reports change in a form that is honest about what was trained and what was not.
A note on language: throughout this page we describe the practice modules in their published form (which is clinical), and we describe the StrokeVoice product itself as a wellness and practice companion. The product is not a medical device, does not diagnose, and does not claim a treatment effect.
Protocol mapping
How each module is grounded.
Each daily practice module in StrokeVoice is informed by a published treatment approach. The list below isn’t exhaustive — it summarises the primary literature we drew on and continue to follow.
In the product today
When semantic cueing doesn’t resolve, the cue ladder moves to phonological support — first sound, then a full spoken model. Complementary to SFA on the same target; the event record captures which route produced the word.
Leonard, Rochon & Laird, Aphasiology, 2008.A five-stage progression — Listen → Read-along → With-cues → Solo → Generalise — persisted per stage so the user and clinician can see fluency consolidation across sessions. Scripts are personalised where the user’s own situations have been captured.
Youmans, Holland, Muñoz & Bourgeois, Aphasiology, 2005. Cherney, Topics in Stroke Rehabilitation, 2010.On the roadmap — scoped but not shipped
An elaboration-based session type where the conversational partner echoes and expands the user’s own production, rather than cueing towards a target. Near-term build on top of the existing conversation partner.
Kearns, Clinical Aphasiology, 1985. Wambaugh & Martinez, Aphasiology, 2000.Requires a verb-centred stimulus set curated with an SLT; we are not claiming VNeST until that stimulus set is in place and reviewed.
Edmonds, Nadeau & Kiran, Aphasiology, 2009. Edmonds, Mammino & Ojeda, AJSLP, 2014.Deliberately out of scope for a self-directed app
We have considered but are not claiming Melodic Intonation Therapy (benefits from live clinical supervision and case-selection), Multi-Modality Aphasia Therapy (requires drawing, writing and gesture inputs that are real product work, not a copy change), or the intensity protocols of CIAT/CILT (which assume daily clinician-supervised dose). We point users towards a qualified SLT for these.
Measurement framework
What we report, and what we deliberately don’t.
Four layers, deliberately separated: passive telemetry on every session, an untrained probe on a fortnightly cadence, a Voice Journal entry monthly, and a confidence and carer check-in available any time. Trained material is kept separate from untrained probes by design. There is no single composite “aphasia score”.
Quiet telemetry
Accuracy, latency, self-corrections, cue-level — captured silently behind normal practice. Stored per item; powers adaptive difficulty. Never shown as a score.
Untrained probe
A held-out probe set, kept separate in the database from trained items. The cleanest read on whether change is generalising; gated by a Minimum Detectable Change threshold before any “you’ve improved” framing is shown.
Voice Journal entry
A short connected-speech sample with the same prompts each time, so they are directly comparable. Each entry is preserved for side-by-side playback against any earlier sample.
Check-in
CCRSA-informed confidence rating, optional CETI-informed carer rating, and a place to log functional wins across twelve everyday categories (family, phone call, café order, shopping, medical, community, hobby, reading, work, celebration, self-care, other).
A note on instruments and cadence: connected-speech metrics like CIU (Nicholas & Brookshire, 1993) are typically computed on longer samples — which is why the Voice Journal runs monthly by default rather than weekly. A 4-week rolling-mean dashboard view of WPM, MLU, TTR and CIU is on the near-term roadmap; until then, each sample is available with its audio for per-sample review. Cadence is configurable for research contexts.
Objective signals
- Probe accuracyNaming accuracy on a held-out, untrained probe set, re-tested fortnightly.
- Trained-item accuracyReported separately from probe accuracy. Improvement here without probe gain is reported as such.
- Response latencyTime from prompt onset to first verbal attempt, tracked per item and aggregated cautiously.
- Self-correction rateProportion of attempts where the speaker repairs without external cueing.
- WPMWords per minute on the Voice Journal sample.
- MLUMean length of utterance in connected speech.
- TTRType-token ratio across the sample.
- CIUCorrect information units per minute (Nicholas & Brookshire, 1993) — informativeness rather than raw output. Rolling-mean dashboard presentation is on the near-term roadmap.
Lived-experience signals
- CCRSA-informedBrief user check on communicative confidence in real situations (Communication Confidence Rating Scale for Aphasia, Babbitt & Cherney, 2010). Available any time; MDC-gated before framing change as progress.
- CETI-informedCarer-rated everyday communication observation (Communicative Effectiveness Index, Lomas et al., 1989). Optional, shares the same MDC treatment.
- Functional winsThe conversations the user logs as personally meaningful — captured qualitatively.
- Wellbeing & fatigueLightweight mood signal so a difficult day is contextualised, not penalised.
- Voice JournalSide-by-side playback of today’s voice and earlier recordings — qualitative, but central to honest review.
Built-in invariants
- Trained vs untrained items are kept separate in the database, so practice cannot contaminate a probe.
- No composite “aphasia score.” Trained accuracy, untrained probe accuracy, connected-speech metrics and self/carer-rated signals are reported alongside each other, never collapsed into a single number.
- No comparison to other users. The only meaningful comparison is to the user’s own previous self.
- Plateaus are reported honestly. A flat stretch is shown as a flat stretch, not dressed up as progress.
Dose & distribution
Why daily, repeatable practice is the lever we’re pulling.
The strongest available evidence on dose comes from the RELEASE collaboration meta-analysis (Brady, Ali, VandenBerg, Williams, Williams, Abo, Becker et al., Stroke, 2022; n = 959 across 25 trials). It associates the largest gains in language outcomes with cumulative practice in roughly the 20–50 hour range, distributed across an episode of care, with intensity and total dose both contributing.
Direct therapy time alone rarely reaches that band in routine UK community practice. StrokeVoice exists to make the time between sessions count, in a structured and recordable way — not to replace direct therapy.
Privacy & data handling
Voice recordings are encrypted at rest using per-user envelope encryption (AES-256-GCM data keys, wrapped with a master key). Recordings are never used for advertising and are never sold. Model improvement from user voice is not enabled by default; a granular, revocable opt-in — with in-app withdrawal that deletes material collected under that consent — is planned for when we begin any such work. UK hosting for UK user data is part of the production deployment plan and will be documented in the Trust & privacy pack.
Get the evidence pack
A more detailed PDF pack — covering protocol implementation notes, measurement schedule, safety guardrails on the conversational AI, and known limitations — is available to clinicians on request.