Vocable
A voice-first vocabulary app — say a word you don't know, and it becomes a spaced-repetition flashcard seconds later.
Spaced repetition is one of the most well-established results in how memory works, and it’s the proven way to build vocabulary. The problem is the tool. Anki does it, but making a single card is a chore — look the word up, copy the definition, fill in a form, pick a deck. That friction is enough to stop most people, and it’s worst for the ones who’d gain the most: students and kids, who aren’t going to fight an app to save a word.
Vocable is the low-friction version. You give it a word and the sentence you saw it in, it finds the definition, and that’s the card — a few seconds, not thirty. Then the proven part takes over: spaced repetition brings each word back right before it would be forgotten, so it sticks. None of the method is new. The only thing I changed is the step that was keeping people out.
How it’s built
The interesting problem was capture. Recognising a word isn’t a transcription problem — it’s a ranking one. Whisper mishears a rare word you’re saying out loud for the first time, maybe one you’ve only ever read, so instead of trusting the transcript Vocable ranks dictionary candidates phonetically and hands you three to tap. The sentence you give it does triple duty: it sharpens the recognition, it picks which sense of the word you meant, and it’s the context that makes the word memorable.
Underneath: Double Metaphone phonetic matching over de-spaced token windows, fuzzy trigram retrieval, embeddings for sense disambiguation, and FSRS — a modern spaced-repetition scheduler — deciding when each card comes back, so I never set an interval by hand. It’s a React PWA that keeps working offline, on a FastAPI backend with Postgres and pgvector. It started local-first, everything running on your own machine, and now also runs as an invite-only cloud beta on free tiers — Cloudflare Pages, Cloud Run, Neon, Groq, Nomic. The right word lands in the top three about 94% of the time.