How Speak built an AI English tutor and became Korea's top education app

Executive overview

Speaking a second language is nearly impossible to learn without expensive human tutors or immersion. Speak replaced the tutor with AI-powered speech recognition, making fluency practice accessible and affordable at scale.

The founders spent a year doing deep AI research before building, then iterated relentlessly on product-market fit by going in-person to Korea to watch real users. Retaining over 50% of subscribers on day 30 proves the model works.

The core insight: pick one market, go there in person, and watch people use the product — then fix what you see.

Founders and origin

• Connor sold his first app, Flashcards Plus, at 21; the exit gave him freedom to pursue passion over money
• Andrew earned three degrees by age 12 and spent 3.5 years of a Stanford neuroscience PhD before dropping out
• Both joined the Thiel Fellowship, became roommates, and spent a year doing deep ML research before building Speak
• Early focus: speech recognition that understood not just words but accents — achieved state-of-the-art results using unlabelled YouTube data

Finding product-market fit

• Launched in every global market initially; nothing converted well enough — users liked the product but didn't love it
• Decided to pick one market and commit: flew to Korea, Japan, and Europe to interview users in person
• Korea was chosen — high English-learning spend (at one point 1% of GDP), dense competition, and very opinionated users
• Early user sessions revealed a simple gap: users wanted to speak more; existing product had too little speaking practice
• First day in the Korean App Store: $18 in revenue — and they celebrated

The commute insight that unlocked growth

• Biggest retention blocker: users wanted to practise on buses and subways but couldn't speak aloud in public
• Counter-intuitive fix: built a silent reading/listening mode for commutes to create daily habit formation
• Once the habit was established, users returned to active speaking practice at home
• Result: conversion, retention, and every other metric spiked immediately after launching the commute mode

Product, ML, and content flywheel

• Early versions used off-the-shelf speech recognition — good enough to launch and collect training data
• That data fine-tuned their custom models, improving accuracy and enabling better product features over time
• Now have an internal ML team building conversational AI features powered by modern language models
• Content is treated as a product: A/B tested and iterated continuously, not built once and left alone
• Marketing: "AI tutor" messaging confused users; "speak 100 sentences in 20 minutes" was concrete and converted
• Key retention metric: over 50% of subscribers are still active on day 30 after starting

Expansion and long-term vision

• After proving Korea, expanding to Japan and the US — applying the same local-first, in-person research playbook
• Ultimate goal: make fluency practice as accessible as any app, at any price point, in any language
• Technology vision extends beyond language: the voice-AI infrastructure they are building applies to any domain where humans communicate with machines