Cognitive decline detection without transcribing voice

A senior's transcript is an intimate document. You can read who they loved, what they regret, how they think. Storing it betrays a moral contract, even with signed consent.

So we don't store it. The transcript lives in memory for 30 seconds — long enough to extract features (syllable rate, lexical richness, hesitation ratio, content/function-word ratio, ...) — then it's gone. Only the features persist, as daily-aggregated time series.

Seven indicators are enough to spot a six-month drift. The secret isn't in the model — it's in signal regularity. When you talk for 5 minutes a day to the same voice, variations stand out.