How Learna Scaled Voice Learning with ElevenLabs

Learna is an AI language learning app centered on real-time voice conversation, and a rapidly growing core product of Codeway. Within 18 months of launch, it recorded 45 million users, 7 million MAU, and 90 million ARR, while the parent company's cumulative scale reached 500 million+ users and $400 million+ ARR.

As scale grew, the problem became clear. When the tutor's voice sounded robotic, beginner users hesitated to speak, lost confidence, and eventually churned. In language learning, voice wasn't just a feature—it was the core product layer that determined whether users would stay.

Multilingual support was also challenging. Learna covers 80+ languages, and needed to handle both a user's native language and their target learning language within a single session. For example, the system had to understand Turkish while simultaneously pronouncing English accurately.

Learna ran an A/B test comparing its existing voice system against ElevenLabs, and the results were statistically significant.

• Trial conversion up 9%
• Customer lifetime value up 10%

With these results, voice was redefined not as a UX improvement but as a growth lever. In particular, the bilingual tutor was the turning point. Once the tutor could naturally switch between a user's native language and target language within the same session, both beginner engagement and monetization improved together.

The practical lessons Learna left behind are clear.

• Treat voice as a product layer, not a feature: Voice quality directly determines trust, session length, and retention.
• Support the native language first: English alone isn't enough to drive sufficient engagement in many markets.
• Validate with business metrics rather than subjective feedback: Confirming impact through conversion rate and LTV builds a stronger case.
• Actively adjust prompts to reduce hallucination: Parameter changes affect pronunciation, reasoning, and consistency, requiring repeated tuning.
• Leverage settings before writing custom code: Language switching, supported languages, LLM assignment, and response timing can all be adjusted through in-platform settings.

Ultimately, this case shows that natural voice can move beyond the felt quality of an AI language learning service to affect conversion rates and revenue as well.