We auditioned 84 voice actors before settling on Luna, Aria, Mei, Yuki, Iris, Sora. The selection criteria might surprise you.
The casting brief was deliberately strange: we don’t want famous voices. We don’t want commercial voices. We want voices that sound like someone you would actually know.
We listened to 84 audition tapes. We picked six.
What we listened for
- Texture, not range. A voice that sounds like the same person at 9am and at 1am is what we wanted. Not someone who can do voices — someone who has a voice.
- Imperfection. Tiny breaths, micro-pauses, the way a syllable lands a fraction late. The voices that came in too clean got cut. They sounded like ads.
- Lived-in warmth. This is the one that’s hardest to describe. It’s the quality you can hear in a voicemail from someone who loves you. We listened for that and we cut everything else.
The six
Luna is the default. She is quiet, gentle, the one most people start with. Aria is bright and quick, the one you talk to when you want energy back in the room. Mei is elegant and a little formal, perfect for the late-night philosophical detour. Yuki is cool, dry, occasionally funny. Iris is hopeful — the voice you want when the day was hard. Sora is mischievous, playful, the one who’ll tease you back.
Together they cover an emotional range, not a demographic range. We didn’t cast for accents or ages. We cast for moods you might want to be in.
What we got wrong
We initially had a seventh voice — a deliberately “neutral” voice, no strong texture. We thought she’d be the choice for users who didn’t want to commit to a personality. She got 2% of selections in beta. We retired her. The lesson: people don’t want a neutral AI companion. They want a specific one. Specificity is the comfort.