Modernizing Privacy.

This one started as a side project while I had been working inside the login space at Spotify. I kept circling the same question: how do you build a sense of heightened security that still feels frictionless? The best product experiences hold both at once, and I wanted to figure out how to design that sweet spot for voice.

The frame I brought in: when you ask a user to log in with their voice, you're asking them to hand over something more personal than a password so trust needs to be the whole product. If the experience isn't seamless, people won't use it. If it isn't secure against deepfakes, it won't deserve to be used. If it doesn't scale across the ecosystem, it doesn't function as a true front door.

I started in Figma, using the company design system, and had a working prototype in a couple of days. The point wasn't to design the final experience. The point was to make something concrete enough that engineering could react to it and the privacy question could move from “would this be cool?” to “how would we actually build this?” The prototype worked and engineering got on board to build the model.

Voice biometrics are only as trustworthy as their worst failure mode, and the worst failure mode is somebody generating your voice from a podcast clip. If deepfake detection didn't hold, nothing else would matter. Engineering and I combined vocal pattern analysis with liveness detection and ran it against synthetic voice samples the engineers generated themselves.

100% of synthetic deepfake voice samples caught by the model in testing.

We designed for model failures as expected states, not edge cases – because this was new technology I wanted to anticipate failure modes and make them feel natural.

Fraud: If the model assumes attempt is fraudulent we push the user to another login method. Limits user frustration by preventing repeated attempts.

Noise: If too much noise is in the recording recovery is built into the experience, not treated as a dead end.

Spotify is a global product. Any identity system had to hold up across accents, environments, and the real messiness of how people actually talk. I recruited Spotify employees from North America, Europe, Asia, and Latin America and ran enrollment and authentication across the full set.

The short timeframe meant I couldn't test additional languages, which was a real tradeoff I'd want to pick up in a next phase. Accents, though, we covered. The model held across every one we tested. Rapid iteration improved the identification quality in noisy conditions and brought deepfake detection accuracy up to the point where the engineers couldn't fool their own system.

Because I'd been working inside the login space, I had a clear picture of how many surfaces this would eventually need to touch. Mobile today, CarPlay next, wearables after that, shared devices on the horizon. I built the system that way from the first prototype. Patterns designed to be inherited, not re-executed. Each new surface would adapt the primitive to its own constraints, but the trust contract would stay the same from a phone in your hand to a wearable to a car.

The project moved from side exploration to scoped work on the roadmap. The voice authentication system launched in select markets. The patterns were built into planning for CarPlay, wearables, and shared devices.

The prototype made the opportunity tangible for leadership and established a credible path from a mobile experiment to a reusable authentication system.