Webinar Recap: How Learna Scales Voice Learning with ElevenLabs

Voice is the moment of truth in language learning. When it sounds real, users keep coming back. When it doesn't, they leave.

This post recaps "Behind the Agent: How Learna Scales Voice Learning with ElevenLabs," - a conversation about how getting voice right became a direct driver of retention, conversion, and revenue for Learna.

Why Voice Quality Matters

A use case like language learning only works if it feels real. That insight sits at the center of how Codeway built Learna, an AI-powered language learning app designed around real-time spoken conversation from day one.

Codeway is one of Europe's largest consumer app companies, with over 500 million users worldwide and more than $400 million in annual recurring revenue.

Learna is their fastest-growing product, reaching 45 million users, 7 million monthly active users, and $90 million in ARR in just 18 months.

But growth at that scale exposed a hard constraint. When the tutor sounds robotic, users don't feel comfortable speaking. They lose confidence. They churn. Voice wasn't a feature they could iterate on slowly, it was the core product layer that everything else depended on.

The specific challenge was around a multilingual offering. Learna supports over 80 languages. Most users are beginners. That means the tutor has to speak in the user's native language while teaching them a target language inside the same session.

A single conversation might require the system to understand Turkish and correctly pronounce English within the same flow.

Behind the Agent: Learna

Learna ran A/B tests in production comparing their previous voice system to ElevenLabs.

The results were statistically significant:

1. Trial conversion increased by 9%
2. Customer lifetime value increased by 10%

These numbers changed how Codeway thought about voice. It was no longer a UX improvement. It was a growth lever.

Their bilingual tutor was the turning point.

Before Learna introduced a bilingual experience, the product showed weaker engagement metrics, especially for beginner-level users.

Once the tutor could speak the user's native language and the target language in the same session, both monetization and engagement improved significantly.

Best Practices from Learna

1. Treat voice as a product layer, not a feature. If you are building a conversation-based product, voice quality affects trust, session length, and whether users come back. Evaluate it with the same rigor you apply to core product decisions.
2. Support the native language first. For most markets outside of major international hubs, users want to interact in their native language. Launching with English only leaves significant engagement on the table. At minimum, pair English with the dominant local language for each market.
3. Test voice quality against business metrics, not just user feedback. Learna did not stop at qualitative signals. They ran A/B tests in production and tied voice quality directly to trial conversion and LTV. That data gave the team the confidence to build the core experience on top of ElevenLabs.
4. Iterate aggressively on prompting to reduce hallucinations. AI systems do not behave like deterministic software. Every parameter change can affect pronunciation, reasoning, and output consistency in ways that are hard to predict. The prompting guide in ElevenLabs documentation is a good starting point. LLM selection also matters. ElevenLabs is LLM agnostic, which gives teams flexibility to tune behavior per language and per use case.
5. Use configuration before you use custom code. Language switching behavior, supported languages, LLM assignment per agent, and response timing are all configurable inside the ElevenLabs platform without requiring custom engineering. Start there before building custom architecture.

Watch the full session

Watch the full session here.