Helping Duolingo Users Build Real Speaking Confidence
Research into why Duolingo learners keep their streaks alive but still cannot hold a conversation
Phase 01 · Problem discovery
Helping Duolingo Users Build Real Speaking Confidence
✳The problem they went after
Duolingo's bite-sized lessons and streak mechanics get people opening the app daily, but what they learn does not travel into spoken conversation. Across 19 survey respondents and 9 interviewees, learners said they felt underprepared to speak, could not tell whether they were being taught formal or casual language, and never had their pronunciation corrected.
✳What they actually did
The analysis lands on sentence formation as the core failure: learners cannot build sentences independently, confidence drains, and drop-off follows in later learning stages. A SWOT analysis and PIF prioritisation scored this gap above notification fatigue and lesson repetitiveness. The proposed direction is interactive, context-rich practice tied to real situations such as travel and business communication, so in-app activity lines up with the goal learners actually hold.
✳What you can take from it
Worth stealing
- 19 survey responses and 9 interviews behind the claim
- PIF prioritisation ranks three competing problems
- Streak engagement separated from actual language gain
- Formal versus casual confusion surfaced in interviews
Read this if
Learning to argue that a healthy engagement metric is hiding a real problem
The transferable bitStreaks measure platform activity, not language acquisition, and the two have quietly come apart. Reading engagement as evidence of value is a standard trap, and this write-up shows what it takes to catch it: talk to enough users that the distance between the metric and the goal becomes impossible to miss.
✳The submission itself
Open the PDF →Reading on a phone? The embedded viewer is cramped — open the PDF in a new tab instead.
Phase 02 · Feature and metrics
Adding a Speak-to-Form Sentences Challenge to Duolingo
✳The problem they went after
Learners get through reading and vocabulary exercises, then stall the moment they have to form and speak an original sentence. Duolingo's speaking practice is thinner and less frequent than what Memrise and Speak offer, so anyone aiming at real conversation has no structured route there. The confidence gap shows up as weak long-term retention, even among users holding daily streaks.
✳What they actually did
The Speak-to-Form Sentences Challenge sits in the Practice tab and gets pulled in through daily notifications. A user receives a short prompt with seed words, records a spoken sentence, and an AI evaluator scores it across grammar, pronunciation, fluency, word choice, and sentence structure. Five or more challenges without errors earn a shareable Confident Speaker badge and XP. Rollout starts with English-to-Spanish learners and widens as speech model accuracy allows.
✳What you can take from it
Worth stealing
- Seed words scaffold an original spoken sentence
- AI scoring across five named speech dimensions
- Badge rate as primary metric, drop-off as guardrail
- One language pair first, gated on model accuracy
Read this if
Designing gamified practice loops on top of AI evaluation you cannot fully trust
The transferable bitThe reward is scoped to a behaviour worth rewarding, five error-free challenges rather than mere participation. Rollout is then gated on speech model accuracy for one language pair before expanding, which is the honest way to plan around a model you do not yet fully trust.
✳The submission itself
Open the PDF →Reading on a phone? The embedded viewer is cramped — open the PDF in a new tab instead.
