Voice testing with elders.

We test Companion's voice with older adults during product development — in living rooms and facility common areas, with consent, with notes. The sessions are short and structured, and the feedback keeps changing what we ship.

The headline finding is a narrow band. People don't want a voice that sounds robotic. They also don't want one that sounds too human and starts trying to make small talk. They want competence and warmth, in small doses. A device that pays attention, answers, and stops talking.

What surprised us.

Names matter more than we thought. Sevah is hard to say with a partial — the v slides. Companion is universal but felt impersonal to one resident, who renamed hers something shorter and started using it without prompting. We now let the name be set at setup and changed any time, and Companion will answer to either.

Word choice matters more than tone. Would you like lands better than do you want. Take your time lands better than no rush. We rewrote the prompt around a small set of phrasings that test well across sessions, and we removed every filler the model used to lean on — no great, no of course, no apologies it didn't owe.

What we changed.

Three things shipped from the last round. The agent no longer initiates small talk; it greets, answers, and waits. It uses one short sentence by default and expands only when asked. And it never asks a follow-up question when a simple acknowledgment will do.

The next change is a slower default cadence — about fifteen percent slower than the model's natural pace — with the option to speed up per resident. A voice that feels like the help someone actually wants, not the help engineers thought they'd want.