Speak is personalizing language learning with AI

Connor Zwick didn't set out to build another Duolingo clone. Speak, the company he co-founded, has spent years betting that the fastest way to learn a language is to actually talk in it, out loud, with something that listens back and corrects you on the spot. In a new conversation posted on OpenAI's blog, Zwick lays out how that bet has evolved alongside OpenAI's models, moving from scripted drills to something closer to a patient, endlessly available tutor.

The core idea sounds almost too simple: instead of matching multiple-choice vocabulary, learners speak sentences and get immediate feedback on pronunciation, grammar, and word choice, tuned to their specific level. Zwick describes this as personalization not in the marketing sense of the word, where an app just remembers your name, but in the sense that the AI adjusts pacing and difficulty in real time based on how a person actually talks. That's a meaningfully different product than a flashcard app with a chatbot bolted on.

What's notable is how directly tied Speak's roadmap is to OpenAI's own progress. Zwick talks about voice and reasoning capabilities that simply didn't exist a couple of years ago, and how each jump in model quality has let Speak strip away workarounds and hacks that used to compensate for AI that couldn't quite follow a natural conversation. Fewer scripted fallback responses, more genuine back-and-forth. That's the quiet story behind a lot of AI-native startups right now: their product improves in ways they didn't build themselves, simply because the underlying model got better.

Zwick also frames language learning as an unusually good testbed for conversational AI generally, since success is measurable and immediate. A model wobbling on a subtle Spanish subjunctive mistake, or nailing the correction, tells you something concrete about how far voice AI has come. Speak has apparently grown fast on the strength of that feedback loop, particularly in South Korea, where the app has become one of the more prominent AI education tools. Zwick's pitch, in the end, is less about gamification and more about giving people enough real practice that they stop being scared to open their mouth in a new language.