We could give Companion her daughter's voice. We won't.
Voice cloning is good enough now to recreate a family member from a few minutes of audio. In a memory-care room that is not a feature — it's a deception. Where we draw the line, and why.
A few minutes of a person's recorded speech is now enough to clone their voice well enough to fool the people who love them. Somewhere in our roadmap conversations someone always raises the obvious application: a resident with dementia is calmer when she hears her daughter. Her daughter lives four states away. We could make Companion speak in the daughter's voice. The technology is ready. The question is whether we should ever point it at a confused person in 214B, and our answer is no.
The line is consent — and who can give it
Two different people have to consent to a cloned voice, and in eldercare both consents are fragile. The first is the person whose voice it is: cloning the daughter requires her informed, revocable, specific agreement, not a buried checkbox. The second is the person who hears it. A cognitively intact resident can be told this is a synthetic voice modeled on your daughter and decide they like it. A resident with moderate dementia cannot hold that frame. To her, it is not a model of her daughter. It is her daughter. That is not comfort we provided. It is a belief we manufactured.
And the people most likely to be soothed by a familiar voice are exactly the people least able to consent to being deceived by one. The benefit and the harm point at the same population. That is the whole problem in one sentence.
Deception is a clinical risk, not a vibe
It is tempting to wave this off as a philosophy-seminar worry. It isn't. A resident who believes her daughter is in the room has expectations that the device cannot meet and that staff then have to manage:
- Reality disorientation — a familiar voice that knows nothing about today can deepen confusion rather than ease it.
- Grief and abandonment loops — why won't she come see me, she was just talking to me — re-triggered every session.
- Eroded trust in the room — when the illusion slips, the one device meant to be a steady presence becomes a thing that lied.
- Consent that outlives the relationship — voices cloned during good times, still speaking after estrangement, death, or a withdrawn permission nobody propagated.
How we'd gate it if we ever did
We default to a clearly synthetic, warm, neutral voice that never pretends to be anyone. If a family and care team ever made a deliberate, documented case for a familiar voice — say, for a cognitively intact resident who explicitly wants it — we would gate it hard: written, revocable consent from the voice owner; capacity-aware consent from the resident or their proxy; a standing disclosure that this is a synthetic voice; per-resident clinical sign-off; and a kill switch the care team controls, not us. The bar is high on purpose, because the failure mode is a vulnerable person being fooled by us.
So the resident in 214B hears a voice that is kind, patient, and unmistakably its own. It never claims to be her daughter. When her daughter does call — and Companion's job is partly to make that more likely, not to replace it — she'll know the difference, because we made sure there was one.