August 4, 2026english-rpg.fernandooviedo.site

English RPG - Spoken-English AI coach

App for training spoken English with an AI coach that answers out loud and corrects you without breaking the conversation. Real-time generative UI. Every session earns XP, levels you up and unlocks branches of a skill tree.

Live view

English RPG

Most people learning English read and understand well. What they don't do is speak. And the reason is concrete: speaking exposes you. With a human teacher there's embarrassment, scheduling and an hourly rate. With a flashcard app there's no conversation at all. English RPG targets exactly that gap: an English coach you talk to out loud, that answers by speaking, and that corrects you without stopping the conversation. On top of that, a gamified environment built around RPG progression. Every session earns XP, levels you up and unlocks branches of a skill tree.

The real problem it solves

Speaking practice has three blockers, and all three are product problems before they are technical ones:

Embarrassment. Making mistakes in front of a person is costly. In front of a coach that doesn't judge, it isn't.
Friction. Booking a lesson takes more effort than the lesson itself. Here you open a browser and talk.
Drop-off. This is the lethal one. Without a reason to come back tomorrow, the user stops practising by week two.

Conversational AI solves the first two. It does not solve the third - and that's where game design comes in.

Why an RPG and not "a conversation app with AI"

Anyone can wire a chat to an LLM that speaks English in an afternoon. That's not the product. The product is the retention loop:

XP per session - speaking earns a measurable reward instead of a vague sense of having practised.
Levels - progress becomes visible, which is exactly what language learning never gives you in the short term.
Skill tree - it turns "improve my English", a goal too large to act on, into concrete branches you can choose.

The underlying decision: AI is the engine, but progression is the product. Without it, the app is an impressive demo people use three times.

The engineering constraint that drives everything: latency

In a text app, 800 ms of waiting is acceptable. In a spoken conversation, it's a broken conversation. The voice loop is a chain where every link adds delay:

user audio → LLM → Thinking → Tooling call → speech synthesis → audio back

Run that chain serially, waiting for each step to finish, and the reply lands late - the user feels they're talking to a machine. The real work isn't integrating the models; it's overlapping the stages so the reply starts playing before it has finished being generated, and keeping session state off the critical path of the speaking turn. That's the difference between an assistant that works and one that's pleasant to talk to.

Correcting without breaking the conversation

The second design problem: when to correct. A corrector that interrupts on every mistake kills the fluency the app is trying to build. One that never corrects teaches nothing. The product's answer is correction woven into the flow: the coach carries the conversation forward and delivers the correct form inside its own reply, instead of halting the turn for a grammar lesson. The user hears it said properly and keeps talking.

Status

In private beta, with user accounts. Try it at english-rpg.fernandooviedo.site.