Language apps teach you words but never make you speak them
“Three hundred days of Duolingo and I froze ordering a coffee. I know the vocab on flashcards — I've never once produced it out loud in a real sentence under time pressure.”— Hacker News ↗
AI conversation partners exist (Quazel, TalkBits) but they're open-ended generic chat — beginners drown. The gap is drilling: short spoken scenarios constrained to YOUR deck (the 50 words you're actually studying plus survival glue), with a report showing which target words you really produced.
Language learners stuck between flashcards and fluency
Build a speaking-drill partner in Next.js. User pastes their vocab list and picks a scenario (café, directions, small talk); the LLM (Vercel AI SDK) roleplays using only that vocab plus basic glue words, speaking via TTS. User replies by mic — Web Speech API for speech-to-text. After each session show a coverage report: which deck words the user actually said. First 3 steps: (1) vocab deck CRUD + scenario picker, (2) the conversation loop (mic → STT → constrained LLM reply → TTS), (3) the post-session report marking produced vs. missed deck words.