Speakable
Real-time AI personalized English conversation app - Adaptive dialogue learning prototype
Overview
A prototype for real-time English conversation practice in which the AI adapts what it says to the learner’s level and interests as the dialogue unfolds. Built at EnglishAI to validate the user experience of voice-first learning.
Core Problem: Latency Is the Whole Experience
In a text chat, a two-second pause is nothing; in a spoken conversation it is the difference between talking to a person and waiting for a machine. Voice interaction tolerates none of the delay a normal request-response stack accumulates, and the moment the learner feels they are waiting, natural speech stops — precisely the behavior the product exists to produce.
The prototype’s central question was therefore architectural rather than linguistic: what must stream, what can be precomputed, and where the model’s response can begin before the learner has finished speaking.
Key Contributions
- Prototyped the real-time voice conversation system end to end, treating perceived latency as the primary design constraint
- Integrated Hume AI EVI with Gemini 2.5 Flash — pairing an expressive voice layer with a fast reasoning layer
- Implemented personalized conversation context — carrying the learner’s level and interests across turns rather than resetting
- Developed level-adaptive response generation — the AI moderating its own vocabulary and pace to stay just above the learner
Tech Stack
- AI: Hume AI SDK (EVI), Gemini 2.5 Flash
- Frontend: Next.js, React, TypeScript
- UI: Tailwind CSS
- Infrastructure: Vercel (Serverless)