A transparency page, not a marketing page. It separates what published science has shown about speech and cognition from what Recalla has actually built and hypothesized — and it's honest about what hasn't been proven yet.
The idea that speech carries signals about cognition is not ours — it's a decades-old, peer-reviewed body of work. Here's what that literature actually establishes, including where it's cautious.
Fluency, lexical diversity, pause patterns, and semantic content shift in neurodegenerative disease — and those shifts can be measured automatically.
Speech-derived measures help distinguish mild cognitive impairment from healthy ageing. A 2025 meta-analysis of 54 studies found ~80% pooled accuracy for MCI — with wide confidence intervals.
Early-life linguistic ability predicted late-life cognition in the Nun Study; language features predicted future Alzheimer's onset in the Framingham cohort.
In autopsy-confirmed Alzheimer's, connected-speech features changed in step with disease progression — evidence that speech can be a longitudinal marker, not only a one-time classifier.
Standardized acoustic feature sets and open shared-task benchmarks mean methods can be compared rigorously rather than in isolation.
Most published results are cross-sectional group classification (people with a diagnosis vs. without), not within-person tracking over time. Performance varies with task, language, and recording quality, and larger prospective validation is repeatedly called for.
Why a complementary signal could matter — and where a digital approach is being explored in the literature.
This is the line where external science ends and our own contribution begins. We state it as a hypothesis under evaluation — deliberately, because that's what it is.
We hypothesize that frequent, low-friction voice check-ins can produce longitudinal signals that complement — not replace — episodic clinical assessment, and that a person's change from their own baseline is more informative than any single snapshot.
Notice what this is not: a claim that Recalla detects or diagnoses dementia. Whether these signals are clinically useful in real-world settings remains under prospective evaluation. The sections below are honest about how far along that evaluation is.
The honest current state. Where results don't exist yet, we say so rather than implying them.
How we intend to move from feasibility to clinical evidence. These are planned milestones and directions — not commitments or completed results.
Concierge pilot; measure adherence and clinician usefulness.
Structured pilot; first longitudinal data readouts.
Prospective study against reference measures.
Multi-site evaluation in care settings.
Large longitudinal cohort; peer-reviewed publication.
Every reference below was verified to its source — real title, authors, journal, year, and DOI. It's meant to be useful to any reader working on voice and cognition, not only to us. Filter by topic.
No. Recalla is not a diagnostic device and does not detect, diagnose, or rule out any condition. It's designed to produce longitudinal signals that could complement clinical assessment. Any score it shows is illustrative and non-diagnostic.
The foundation is. Decades of peer-reviewed work show that speech and language change with cognition (see the Scientific foundation and Research library). What's ours — and still a hypothesis under evaluation — is the specific bet that frequent, low-friction, longitudinal voice check-ins are clinically useful in the real world.
To date: a working prototype, a feature-extraction pipeline, an explainable scoring framework, and an open technical preview. We have not completed a pilot, performed clinical validation, or published peer-reviewed results. A concierge feasibility pilot is in design. The What we've built section tracks this honestly.
Most prior work is cross-sectional group classification — telling apart people who already have a diagnosis from those who don't, at a single point in time. Recalla focuses instead on within-person change over time, as a complement to episodic testing. Whether that's more useful in practice is exactly what we intend to test.
The interactive demo uses synthetic participants only — no real patient data. Its live check-in features run in your browser and produce modeled, illustrative values from your own voice; nothing is uploaded. Pilots will be run under appropriate consent and ethics review.
Yes, by design. Every composite decomposes into the individual signal families that produced it, so a clinician can see why a value moved. Explainability is increasingly recognized as essential for clinical adoption and regulation of speech-based tools (Shankar 2025).
The full, DOI-linked Research library is on this page. Our public technical framework and the interactive demo are linked from the Resources page. This Evidence Hub will grow as our own results and publications appear.