Korean has small, intimate words. Words you only use with people you’re close enough to that titles would be a wall. Chingu is one of them. It means friend. Add the vocative -ya and you have the way you call out to a friend across the room. Chingu-ya.
The first version of the product had a different wake word. We started with the polite form, the kind a delivery driver uses with a stranger. We tested it with three Korean halmonis (grandmothers) in Daegu, in a session conducted in Gyeongsang dialect by a researcher who has done eldercare work for twelve years.
All three said the polite form felt like the bank phone tree. None of them used it. We asked what they’d say to a real friend they wanted to call across a room. The answers came back fast. Chingu-ya. “Hey friend.” The familiar form. The form a daughter would use with a school friend, or a friend in the village, or someone you’d done morning exercises with for forty years.
We changed the wake word that night.
Why the small word matters.
Building voice AI for older Korean adults is a discipline of register. Korean has formality levels stitched into every verb ending, and the wrong level is not a small mistake. It’s the difference between “please pay your bill” and “hey, dinner’s ready.” If a grandmother calls out to her companion and the companion’s name is in a register she’d use with a clerk at the post office, she will not call again. She’ll keep the phone in the drawer.
The wake word is not a feature. It’s the contract.
We hold ourselves to a few specifics on the contract:
- The wake word is processed on-device. No audio leaves the phone until the senior says Chingu-ya. If she sneezes, hums, plays the radio, none of it is recorded. PIPA Article 23: voice is a biometric, separate consent, full encryption.
- Chingu announces itself. First call, the senior is told this is an AI, not a person. AI Basic Act Article 31: AI identity disclosure on every initiation.
- Voice clips delete automatically. 90 days, unless the senior or the family opts in to keep them. PIPA Article 21.
- Stored in Korea. Not in a US data centre. Not in Frankfurt. The country the senior already lives in.
What “one word” lets us avoid.
A lot of voice products start by demanding a sentence. Wake-word + verb + object. Hey product, please tell me the weather. That’s a syntax exam, not a conversation. With Chingu-ya, the senior says one word and Chingu opens. From there it’s a conversation, in her dialect, at her speed.
We’re collecting dialect corpus across five regions: Gyeongsang, Jeolla, Chungcheong, Gangwon, Jeju. Standard Korean is live; the regional models are training. The expectation is that a halmoni in Daegu speaks Gyeongsang to her companion, not Standard, because that’s how she would speak to a friend.
The night the wake word changed.
After the Daegu session, our researcher sent us a four-minute voice memo. She had taken the train back to Seoul and recorded it from the window seat. I’ll paraphrase the part I keep coming back to:
She said the third halmoni had paused at the new wake word. Held it on her tongue. Then said, chingu-ya, you make it sound like I’m the friend. And smiled. That’s the build.
That’s the build. The companion isn’t the friend. The senior is. Chingu is the seat the friend can sit in until a real friend can come back.
What we’re still working on.
The dialect coverage is incomplete. Five regional models is the goal; right now Standard is solid and Gyeongsang is in test. Jeju, which is so distinct from Standard Korean that some linguists call it a separate language, will take longer.
We are not running paid pilots yet. We are not certified by KISA yet. We have no revenue yet. The wake word works, the conversation works, the alert routing works, the family app works, the welfare dashboard works. The certification stamps and the revenue come from doing the work, not from declaring readiness on a slide.
Three Korean municipalities are in active conversation. One welfare NGO. One Samsung SmartThings partner team. We will name them when there is something concrete to name.
ETB · voice memo from a kitchen in Seoul, February 2026