Every language learner gets a tutor now

The best way to learn a language was always a patient tutor. For the first time, everyone can have one. Here's what that changes for learners, teachers and schools.

The hardest part is speaking

Ask anyone who's learned a second language what the hardest part was. Most will say speaking.

Last month, we built an AI tutor to practice exactly that. Hundreds of thousands of people watched it teach someone to order coffee in Japanese.

https://x.com/wayne_liang_/status/2098119361894752627

A lot of the shares came from native Japanese speakers. One parent said their daughter had given up practicing English with a voice-only AI because there was nothing to look at, and might try again with this. Another begged for Taiwanese Mandarin. A developer just asked: "Now, what should I build?"

Nobody asked whether it looked real enough. They just wanted it. That's what speaking practice has always needed: a patient partner, in any language, whenever you want one. Most learners never had one.

The 40-year-old problem

In 1984, the education researcher Benjamin Bloom published what he called the 2 Sigma Problem. Students with a 1:1 tutor outperformed classroom students by two standard deviations. The average tutored student did better than 98% of the class.

Bloom, B. S. (1984). The 2 Sigma Problem. Educational Researcher, 13(6), 4–16. doi.org/10.3102/0013189X013006004

We've known the best way to learn for more than 40 years. We just couldn't give everyone a tutor.

Language makes it harder. You can't learn to speak from a textbook. You need someone across the table, listening, reacting, waiting while you find the word. Ideally, someone who speaks both your mother tongue and the target language, so they can explain what you got wrong. That person is hard to find for most language pairs, costs money, and keeps office hours. And in front of another person, every mistake feels like it's being watched.

So most learners use what's available: flashcards, apps, grammar drills. Then they freeze the first time a real person talks back.

What changed this year

Three things arrived at once.

Voice models that listen while they talk. You can interrupt, stumble, change your mind mid-sentence, and the conversation keeps going.

A face that reacts in real time. It nods, pauses, and looks confused when you say something strange. That's what real conversation feels like, and what voice alone can't train.

Costs that keep falling. Live practice used to be a luxury feature. Now it's priced by the minute, and the minute keeps getting cheaper.

Our Japanese tutor puts the first two together. We built it as one of the launch partners for OpenAI's GPT-Live, with a LiveAvatar face. Try ordering a coffee and changing your mind halfway through: the tutor keeps up, fixes your pronunciation and teaches you the word you were missing. It feels messy, fast and a little nerve-wracking, just like a real conversation.

The third change is the one we've worked hardest on. When we asked customers what held them back, the answer was scalability with cost-efficiency, not realism. So we spent six months making every live session more efficient to run. The result: unlimited concurrency on every paid plan, and minutes as low as $0.01 from our enterprise tiers. The full story is in our post:

https://x.com/TryLiveAvatar/status/2092281613682123061

Quality decides whether a tutor is good. Cost decides who gets one. You can already see this at the biggest app in the category. Duolingo has called speaking practice its biggest gap. Its video calls with Lily more than doubled the words each learner speaks, and as AI costs fall, it's moving those calls to a cheaper plan so about 10x more learners get access.

That changes things for everyone in language learning: learners, teachers and schools.

For learners: a tutor at 2am who never runs out of patience

Picture someone with an IELTS speaking test on Friday and a visa riding on the score. It's 2am. Their teacher is asleep. The exam is face-to-face with an examiner, and that's exactly the part they're scared of.

Tonight, they can sit a full mock exam. Then do it again.

Or picture someone learning Korean to talk to their partner's grandmother. They have a hundred questions they'd be embarrassed to ask a person. "How do I say that politely?" "Sorry, what was the word for delicious again?" An AI tutor has heard every question and never gets tired of them. Researchers keep finding that practicing with AI lowers speaking anxiety in the moment, and that anxiety is what stops most people from speaking at all.

Any language, any accent, any setting: a café in Tokyo, a job interview in London, dinner with the in-laws. And you pick the tutor you click with: their voice, their personality, how patient they are. Then they pick up right where you left off, every time.

For teachers: more teaching, fewer reps

Ask any language teacher how many times they've asked "So, where are you from?"

A huge share of teaching time goes to reps: the same warm-up questions, the same corrections, the same drills. They're necessary, but they're not what great teachers are great at.

AI can take the reps. Teachers keep the moments that matter: the nuance of a culture, the push when a learner wants to quit, the lesson plan only someone who knows the student could design. Students arrive warmed up with the basics done, so the hour with a human goes further.

Harvard Business School already runs this model. In Foundry, its eight-week program for founders, instructors still teach live every week. Between sessions, the LiveAvatars of those same instructors give every founder feedback on practice pitches and mock board meetings. It isn't a language program, but it's the same split: the teacher teaches, the avatar runs the reps.

The program's director told The New York Times she first pictured something closer to a chatbot. Students asked for a more guided experience instead. One of the instructors, venture capitalist Jeff Bussgang, admitted his digital twin is a little "creepy." His verdict anyway: "My students love it."

For schools and institutes: the part you could never scale

Every language school teaches speaking. Few can give every student enough of it, because speaking practice needs a person on the other side, and a school only has so many teachers.

Exams show the gap most clearly. Many Cambridge exam prep apps have no speaking section at all. Not because speaking doesn't matter. Because it needs a live examiner, and you can't ship a human inside an app.

High-stakes speaking tests put a person across the table. IELTS is face-to-face. OET, the English test for doctors and nurses, is a patient roleplay. Even TOEFL's 2026 redesign added a "Take an Interview" task with a video of a person asking the questions.

Speaking was always the section schools couldn't scale. Now every student can have an examiner every day, and teachers get transcripts and scores instead of guessing who practiced. One teacher can give an entire cohort daily speaking practice. And a student who could never afford a private tutor gets the same practice as one who can.

The shift happens fastest once learners get to try it. One corporate language training company we work with called AI avatars a gimmick. Then they put a demo on their booth at a learning-tech conference in France. The line ran out the door. They now have a client willing to pay more specifically for it.

The skeptics didn't change their minds from a pitch. They changed them watching learners.

If you're building in language learning

We've spent the last few months talking with language companies, from 20-year-old institutes to 2-year-old apps. None of them asked why learners would want this. They asked where to start and how to make the numbers work. Here's what we tell them.

1. Start with your scariest moment, not your whole app. The mock exam, the job interview, the first real conversation. That's where an AI tutor earns its place first. Then put it in front of real learners. Their reactions will convince your team faster than any deck.
2. Bring your existing faces to life. Most language products already have faces in them: slideshows, filmed actors, pre-recorded video. They're just frozen, and expensive to update. You don't need a perfect avatar to start. Learners judge the conversation, not the pixels, and what they notice first is the voice and how fast it answers.
3. Design pricing around usage. Live conversation is billed by the minute, not by the feature. Pick the model that fits your business, whether that's practice credits, exam packs or a premium tier, and plan for the price per minute to keep falling. Teams that wait for it to get cheap enough will find their competitors already shipped.
4. Build for teachers, not around them. Let AI handle the reps and hand the important moments back to humans. That's how teachers end up championing it instead of fearing it.

Links for builders:

• Open source repo: https://github.com/heygen-com/liveavatar-gpt-live-demos
• Step-by-step guide: https://docs.liveavatar.com/docs/guides/gpt-live-language-tutor
• Agent shortcut: npx skills add heygen-com/liveavatar-agent-skills, then ask your coding agent to "run the LiveAvatar language tutor demo with GPT-Live"

Who gets to be fluent

For most of history, fluency belonged to people who could afford a tutor, move abroad, or had time to embarrass themselves in public.

That's changing. Every learner gets a patient partner in any language, at any hour. Every teacher gets time back for the work only they can do. Every school can finally teach speaking, the skill that matters most, at scale.

Somewhere tonight, someone is ordering coffee in Japanese for the tenth time, changing their mind halfway, trying again. Soon they'll order it for real.

Speaking is still the hardest part. Now nobody has to practice it alone.