Back office ops · Production

Duolingo scales DuoRadio from 300 to 15,000+ episodes with generative AI, growing daily sessions 10x in fewer than 6 months

The problem

DuoRadio's reach was constrained by a labor-intensive manual production process: crafting 300 episodes for a handful of courses took nearly a year, requiring meticulous scripting, voice actors, and specialized audio editing, leaving only a small percentage of learners with access to the feature.

First attempt

Two early generative AI attempts both failed: generating original scripts from scratch produced subpar results requiring extensive manual editing, and automated translation of existing English episodes frequently missed translation accuracy and proficiency level targets, again leading to time-consuming revisions.

Workflow diagram · grounded in source
1
Curriculum content input
trigger
“feeding existing content from our learning curriculum delivered far better results as it gave our generative AI model specific patterns to follow”
2
AI script generation
ai_action
“By supplying the prompts with well-crafted sentences and exercises created by our Learning Designers for Duolingo lessons, it was possible to generate a large volume of promising scripts (level-appropriate, grammatically sound, using the…”
3
Exercise placement optimization
ai_action
“we leveraged learner session data to place exercises optimally within episodes. Standardizing exercise order and placement helped keep the session structure consistent and made automation more reliable.”
4
AI quality filtering
validation
“we created a bunch of extra episodes and built a generative AI-powered filtering process that assessed scripts on naturalness, grammaticality, coherence, logic, etc.—so only the best made it to learners”
5
Evaluator prompt refinement
feedback_loop
“Our Learning Designers refined these evaluator prompts over time to filter for better and better content. With each step along the way, they raised the bar for our learners.”
6
Automated TTS audio production
ai_action
“Advanced Text-to-Speech (TTS) enabled us to automatically create lifelike voiceovers in multiple languages. Meanwhile, audio hashing (a technique for quickly storing and retrieving pre-generated audio) ensured consistent intros and outro…”
7
End-to-end pipeline deployment
output
“we built end-to-end pipelines to automate the entire lifecycle, from script creation to final deployment. With zero human intervention post-initiation, DuoRadio scaled at a pace once thought impossible.”
Reported outcome

DuoRadio daily sessions grew from 500K to 5M in fewer than 6 months; total episodes expanded from 300 to 15,000+ and courses from 2 to 25+ in fewer than two quarters — a task that would otherwise have taken 5+ years — while saving 99% of costs.

Reported metrics
DuoRadio daily sessions500K to 5M
Time to grow daily sessions 10xfewer than 6 months
total DuoRadio episodes300 to 15,000+
Courses covered2 to 25+
Show all 7 reported metrics
DuoRadio daily sessions500K to 5M
time to grow daily sessions 10xfewer than 6 months
total DuoRadio episodes300 to 15,000+
courses covered2 to 25+
content delivery time (automation vs manual)5+ years reduced to fewer than two quarters
cost savings from automation99%
initial manual production time for 300 episodesnearly a year
Reported stack
generative AIWorkflow BuilderText-to-Speech (TTS)audio hashing
Source
https://blog.duolingo.com/scaling-duoradio/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DuoRadio daily sessions grew from 500K to 5M in fewer than 6 months; total episodes expanded from 300 to 15,000+ and courses from 2 to 25+ in fewer than two quarters — a task that would otherwise have taken 5+ years —…

What tools did this team use?

generative AI, Workflow Builder, Text-to-Speech (TTS), audio hashing.

What results were reported?

DuoRadio daily sessions: 500K to 5M; Time to grow daily sessions 10x: fewer than 6 months; total DuoRadio episodes: 300 to 15,000+; Courses covered: 2 to 25+ (source-reported, not independently verified).

What failed first in this deployment?

Two early generative AI attempts both failed: generating original scripts from scratch produced subpar results requiring extensive manual editing, and automated translation of existing English episodes frequently miss…

How is this back office ops AI workflow structured?

Curriculum content input → AI script generation → Exercise placement optimization → AI quality filtering → Evaluator prompt refinement → Automated TTS audio production → End-to-end pipeline deployment.

Related back office ops cases