Bilingual people don’t switch languages neatly. They mix them in one sentence, borrow a word when it fits better, and change their minds halfway through. Most voice assistants expect one language at a time.
How people talk also changes with the device. Asking a phone, a watch or a car for the same thing sounds different each time.
This work is under NDA, so the process, designs, and results aren't shown here. I'm happy to talk through what I can in conversation.
Working on voice assistants taught me that people don’t adapt to technology, technology has to adapt to them. Bilingual speakers mix languages, correct themselves and shift how they talk from one device to the next. Good design starts by listening to that, not smoothing it out.
As a bilingual speaker myself, this work felt personal. It changed how I think about AI: understanding someone isn’t only about getting their words right, but about recognizing how they naturally express themselves.
People don’t talk to AI the way a system expects. The messy parts are the real parts.
People speak differently to a phone, a watch or a car. Context shapes the words.
Translation alone isn’t enough. It has to sound like the people it’s for.