Case Study: Duolingo achieves scalable, high-quality pronunciation audio with Amazon Polly

Powering Language Learning on Duolingo with Amazon Polly

Duolingo, the world’s most popular language‑learning platform with over 170 million users, faced the challenge of delivering accurate, natural‑sounding pronunciation audio at scale while keeping costs and iteration time low. Recording human voices was slow, expensive, and hard to maintain across dozens of languages, so Duolingo evaluated text‑to‑speech and adopted Amazon Polly to meet the needs for fast, scalable, and controllable audio generation.

Duolingo built a TTS microservice (using Elastic Beanstalk, DynamoDB, SQS, S3, and CloudFront) that pregenerates, caches, and serves Amazon Polly audio globally, and ran six A/B tests comparing Polly voices to their prior TTS provider. Amazon Polly voices (Sally, Carla, Vitoria, Hans, Astrid, and Ruben) won those experiments, produced better user feedback on naturalness and pronunciation, reduced costs and iteration time, and are now used for most languages—delivering measurable, statistically significant improvements in engagement and learning outcomes.

Duolingo

André Kenji Horie

Software Engineer