Verse Catch
App that detects Bible verses in sermon audio using AI (Whisper, Gemini) and real-time updates.
- Type
- Web
- Domain
- AI / Audio
- Status
- Live
Overview
Verse Catch listens to a sermon or talk and shows the Bible verses being referenced as they are spoken.
Audio is recorded in the browser in short chunks, transcribed with OpenAI Whisper, and stitched back into complete sentences before Google Gemini looks for explicit, implicit, and contextual verse references.
Every reference Gemini returns is checked against a verse database across several translations, so a verse the model invents never reaches the screen.
Detected verses are pushed to the browser in real time with Pusher. The database runs on SQLite in development and PostgreSQL in production.
The constraint
Speech is continuous, but transcription works on separate chunks of audio, so a single verse reference can be split across two of them. And the verses shown have to be real, even though they are detected by a language model that can invent references.
What I owned
- Recorded audio in the browser in short chunks and converted each one to WAV before upload
- Transcribed each chunk with OpenAI Whisper, tagged with a sequence number so the text stays in order
- Built an accumulator that buffers text until it forms complete sentences, and resets when a chunk arrives out of sequence
- Detected explicit, implicit, and contextual references ("the next verse") with Google Gemini, keeping only matches above a confidence threshold
- Checked every detected reference against a verse database, falling back to NIV when a translation is missing, so invented verses never reach the screen
- Pushed detected verses to the browser in real time with Pusher
- Ran SQLite in development and PostgreSQL on Neon in production, with a warmup endpoint that keeps the database ready
Stack
- Next.js
- TypeScript
- Prisma
- PostgreSQL
- OpenAI Whisper
- Gemini
- Pusher