Real-time Bible verse detection for live sermons, packaged as a macOS desktop app.
American Missionary Church AI listens to the sermon audio feed, transcribes it locally with Whisper, and detects Bible verses three ways at once:
- Address — "John three sixteen", "turn to Romans chapter eight"
- Quote — the preacher reads the verse text
- Paraphrase — "nothing can separate us from God's love" surfaces Romans 8:38-39
Detected verses can be sent to the projection or broadcast system as an NDI video layer, which ProPresenter, vMix, and OBS all accept as a source. A Stream Deck or any HTTP/OSC controller can drive the app remotely.
Everything runs on the machine. No API keys, no per-Sunday cost, no internet required in the sanctuary. Deepgram cloud transcription is available as an optional alternative.
Signed and notarized builds are on the Releases page.
Open the DMG, drag the app to Applications. Apple Silicon, macOS 12 or newer; the Apple
Intelligence lane needs macOS 26. See RELEASE.md for how releases are cut.
A full-window setup assistant opens the first time the app starts: Welcome, Speech, Detection, Microphone, Ready. It asks for the speech engine (local Whisper, recommended, or Deepgram cloud with an API key), which Whisper model to use, and which microphone to listen to. Run it again from Settings, Help.
Whisper models. The app ships large-v3-turbo (8-bit). The Speech step and Settings, Speech
Recognition can download others from the whisper.cpp repository into the app's data folder:
full large-v3 (3.1 GB, most accurate), large-v3 5-bit (1.1 GB), and full-precision turbo (1.6 GB).
Any whisper.cpp .bin file can also be chosen by hand.
Every finished phrase from the transcriber goes through up to four lanes:
- Spoken references. An automaton of all 66 book names plus a number parser. Instant. A chapter without a verse ("turn to Ezekiel chapter 9") is shown at once and refined when a verse is said.
- Read-aloud quotes. Keyword search against the verse text in the SQLite database. Results appear in Recent Detections for the operator.
- Reading mode. Once a chapter is open, follows the reader verse by verse by word overlap.
- Apple Intelligence (optional, macOS 26+). Fragments the reference parser could not resolve
are sent to Apple's on-device language model through a small Swift helper
(
src-tauri/sidecar/apple-intelligence/). It recognises quotes and allusions such as "the prodigal son" in one to three seconds. Claimed quotes are checked against the real verse text before they are shown. Off by default; switch it on in setup or Settings, Verse Detection, where a "Try it" box shows what the model answers and how long it took. Nothing leaves the Mac.
Speed. The Whisper model is loaded once at app launch and kept resident, so Start is instant. Decoding uses Metal flash attention and a 15-second encoder window instead of Whisper's default 30 seconds, since phrases are at most 10 seconds. Phrases split by a short pause are stitched back together before decoding, and nothing under a second is discarded. The transcript panel shows a live status chip: Listening, Hearing you, Transcribing, Finding verse.
None of this needs an internet connection or an API key. Deepgram is the only cloud component and only if chosen.
bun install
bun run setup:all # Bible database, embedding model, verse index, Whisper model (one-time, ~4 GB)
bun run tauri build # produces src-tauri/target/release/bundle/macos/American Missionary Church AI.app and a .dmg
Requirements: Bun, Rust, CMake (brew install cmake), Xcode 26 or newer (the Apple Intelligence helper is compiled with swiftc against the macOS 26 SDK), Python 3.9+.
The build bundles the Bible database, embedding model, verse index, and Whisper model into the
.app, so the resulting app is self-contained.
bun run tauri dev
American Missionary Church AI is a branded build of Rhema by OpenBezal,
released under the MIT License. See LICENSE.