Founded by former Thiel fellows, AI language tutoring app Speak started its journey in Seoul, South Korea. But as it enters the U.S. market, it’ll have to compete with better-known rivals like Duolingo.
On a trip to Seoul in 2018, Connor Zwick drove past what looked like normal skyscrapers. But he soon learned they were filled with classrooms, dedicated to teaching English, and ads for such classes were plastered across taxi tops and billboards. It hit him: The biggest market for his then-nascent language learning app Speak was here, not in Silicon Valley, where his headquarters were located. “English language learning there was like an obsession,” he recalls. “There was such latent demand.”
But the language training on offer was often ineffective. Students spent hours learning the basics from textbooks and pre-recorded videos, overseen by instructors who lacked fluency. The antiquated methods emphasized learning grammar and vocabulary over actually speaking out loud. “Everything was so academic and there was this fear of making mistakes,” Zwick said.
For Zwick, Seoul’s bustling downtown and its endless demand for English language instruction was both inspiration and opportunity— to build a true AI tutor that could replace the human language tutor and offer a judgement-free space for learners to make and correct mistakes while speaking another language. The promise: “No one’s going to know if you said something embarrassing,” he said.
Today, Speak has a voice-based AI coach that role plays scenarios like ordering drinks at a restaurant, asking for directions to a tourist spot or making small talk with a classmate to help practice speaking a new language in real world situations. Built primarily on OpenAI’s models, the goal is to encourage users to practice enunciating commonly used words and phrases out loud. Learners can also create their own situations from scratch by prompting its AI software. Features like leaderboards and streaks that track daily usage help with consistency. And it’s not just for English-learners. Speak offers its role plays and lesson plans for five other languages: Korean, Spanish, Japanese, French and Italian.
Today some 15 million people have downloaded Speak to practice conversing with its interactive AI tutor and become more fluent and confident speaking in a foreign tongue. That traction has helped the $1 billion-valued startup reach a recent milestone— the company announced today that it has more than $100 million in annualized revenue. It’s largely thanks to the strength of its consumer business. Users start free on the app and pay anywhere from $80 to $200 for access to more content.
But Speak also began pushing into enterprise in 2024, as some of its consumers asked employers to cover its costs. Now, some 500 companies including KPMG and HD Hyundai offer Speak subscriptions to employees primarily in South Korea. And after gaining ground in other countries like Japan and Taiwan, Speak started making a concerted push in the U.S. market earlier this year in June.
That’s impressive progress, but there’s a long way to go to catch up with its biggest rival. Speak’s revenue is dwarfed by that of Duolingo, which booked $724 million in sales last year and projects its business will balloon to $1.02 billion by the end of this year. It’s betting big on AI as well and now offers casual language learning video calls with its AI character, Lily delivering on billionaire CEO Luis Von Ahn’s mission to spin up an automated tutor that can easily and affordably teach anyone a foreign language.
Zwick’s keenly aware of Duolingo’s dominance, especially in the U.S., but believes Speak is well differentiated. Duolingo is geared toward improving grammar and vocabulary, he argues. Speak’s focus is on conversational fluency, building language muscle memory and refining pronunciation so people become comfortable speaking aloud. It’s all in the name— Zwick claims people speak five to 10 times more on Speak than on other language learning apps.
And Zwick isn’t shy about how those differences translate into people’s experiences using Duolingo compared to Speak. “Mobile games that also teach you language are, at the end of the day, ways to feel less bad about using your phone when you’re bored. It’s a healthier alternative to Candy Crush, if you will,” he told Forbes. “And I think what we’re trying to do here…is something a little bit more serious.”
Speak has attracted attention from VC heavyweights like Khosla Ventures, Accel and OpenAI Startup Fund, raising some $160 million in total capital. But funraising back in 2016 when Speak was founded wasn’t easy. Not many people cared about language learning nor AI and some investors even asked the founders if the app was a secret ploy to collect and sell data, Zwick said. But his insight around South Korea’s promise was prescient. Accel Partner Ben Quazzo, who flew to the startup’s user conference in Taipei city and led its $78 million Series C round last December, said Speak’s decision to focus on the South Korean market was intentional— a “sandbox” to test the product in one of the most competitive language learning markets. “We’re not selling like a dopamine hit or engagement bait app,” Quazzo said. “We’re selling an education journey to get you to an end outcome.”
Zwick met his cofounder Andrew Hsu through the Thiel fellowship in 2012. He’d just dropped out of Harvard for what he calls a “deep learning sabbatical” and was building his first startup, a study tool-cum-mobile app called Flashcards+, which was later acquired by edtech company Chegg. In Hsu, he found a kindred spirit, a guy who dropped out of his Stanford PhD in 2011 to join the fellowship. Fascinated by AI, the pair snuck into classes at the University of California, Berkeley and Stanford, refining their understanding of core machine learning techniques like reinforcement learning. They even listened in on lectures by OpenAI cofounder John Schulman without ever getting caught.
Today, OpenAI is fundamental to the company’s growth. As AI models continue to improve, so does Speak’s app, which uses the models to automatically update lesson plans based on users’ performance and can do things like accent correction, which wasn’t possible when Speak was founded. Khosla Venture Partner and Managing Director Sven Strohband, who was an early investor, remembers early 2017 versions of Speak which couldn’t yet carry out full fledged conversations with AI but were able to correctly understand foreign accents. “They were AI-first before it was actually cool to be AI-first, to be honest,” Strohband said.