I'm Alperen, an independent developer. For the past months I've been building TalkSmart, an app where English learners practise speaking out loud with an AI voice partner and get a report on pronunciation, fluency and grammar afterwards. It runs on iPhone and in the browser, and the interface comes in 14 languages.
The "talk to an AI" part sounds easy in 2026. The hard parts were all the places where things failed silently. Here are six of them, with what I changed.
The stack, briefly
Mobile: React Native + Expo. Microphone capture and playback with react-native-audio-api (16 kHz PCM in, 24 kHz PCM out).
Web: plain Web Audio API.
Backend: Node.js + Express, one WebSocket per speaking session, deployed on Railway.
Voice: Gemini Live API.β¦