Voice is not a feature: what Learna learned at 45 million users

Learna teaches English by conversation. The thing we got most wrong at the start was treating the voice as a finishing touch.

Learna teaches English by talking. Not drills, not flashcards, not a streak counter — a conversation with an AI tutor that adapts to your level and keeps going when you stumble. In eighteen months it has become the fastest-growing product in the portfolio: forty-five million people have used it, and more than ten million use it every month.

When you build a product whose entire interface is speech, you learn quickly which parts of it are decoration. The voice is not one of them.

The bilingual constraint

The requirement that shaped everything is easy to state and hard to satisfy: a good tutor speaks your language. Not metaphorically. A Turkish speaker learning English needs a tutor that explains in Turkish and demonstrates in English, correctly, inside the same sentence — switching between a language it must pronounce natively and a language it must pronounce as a model.

Learna does that across more than eighty languages. Most speech systems are good at one language and passable at the rest, and the failure mode is not a bad accent — it is a tutor that sounds authoritative in English and slightly wrong in the learner’s own language, which quietly undermines the one thing a teacher needs.

So we set three hard requirements: naturalness, consistency from one language to the next, and latency low enough that a pause feels like thinking rather than buffering. We assessed several providers against them. ElevenLabs was the most consistent across languages and the most responsive in real time.

Measuring a voice against the P&L, not against taste

Here is the part we would do the same way again. Voice quality is the kind of thing everybody has an opinion about and nobody can settle in a meeting. Demos are useless for this: every provider sounds excellent in a demo, because a demo is a sentence somebody chose.

Instead we A/B tested in live production — our existing text-to-speech engine against ElevenLabs, with real learners — and measured the two things that decide whether a product survives: whether people convert to a trial, and what they are worth over their lifetime.

uplift in trial conversion

9%

uplift in lifetime value

11%

languages the tutor speaks

80+

A better voice moved trial conversion by nine percent and lifetime value by eleven. Nothing about the curriculum changed. Nothing about pricing or onboarding changed. The tutor just sounded like a person, and people stayed longer and paid more.

What we would tell the next team

• Treat voice as a product layer with an owner and a roadmap, not as an integration somebody finishes in a sprint.
• Support the learner’s native language before you add another accent in English. Coverage where your users actually are beats polish where they are not.
• Hold the voice to a business metric. If you cannot name the number it should move, you are choosing on taste.
• Budget real time for prompting. Most of what reads as a model problem — rambling, inventing corrections, breaking character — is a prompt that has not been iterated on yet.
• Exhaust the platform’s own configuration before you write code around it. The custom layer you skip is the one you would otherwise maintain forever.

The broader lesson is about where quality compounds. Codeway serves more than six hundred million people across the portfolio, and the components that travel furthest between products are the ones we treated as infrastructure rather than as features. Voice turned out to be one of those. It started as Learna’s problem and became a capability every product can reach for.