Live listening
Microphone audio is captured, split on voice activity and transcribed locally, with the engine and model switchable from the console.
Speech & presentation · Desktop · Offline-first
LumosCast
Spoken references.
Ready for the screen.
A presentation system that listens to a live sermon, detects spoken Bible references, retrieves the passage from a local database, and puts it on the projector or OBS.
- Role
- Architecture and full build
- Platform
- macOS, Windows and Linux
- Network
- Runs with no internet
Core features.
Everything between the preacher's voice and the passage on the screen.
Reference detection
A parser turns messy transcripts into references: misheard book names, spoken numbers, and partial references spread across utterances.
Offline scripture
KJV, ASV and BSB are bundled in SQLite. NIV, AMP and MSG can be added through api.bible and are cached locally.
Operator console
A React console to search, preview, confirm and queue passages before they go live.
Displays for projection and OBS
Each configured display is a web page on the local network, styled per screen and updated instantly.
A desktop launcher
An Avalonia launcher starts the server, lives in the tray, picks the microphone, and recovers the server if it stops.
Architecture
One process, from microphone to screen.
Hover or tap a component to trace its connections.
Desktop launcher
Starts and supervises the server, tray, microphone picker
- Avalonia
- Web host · starts, restarts
Web host
Pipeline, minimal API and the event broadcaster
- ASP.NET Core
- SSE
- Operator console · SSE + REST
- Display pages · SSE /events
Operator console
Search, preview, confirm and queue passages
- React
- Vite
Audio capture
Microphone input cut into speech at pauses
- PortAudio
- VAD
- Speech engine · speech chunks
Speech engine
Local transcription, biased to scripture vocabulary
- Whisper.net
- sherpa-onnx
- Reference parser · transcripts
Reference parser
Finds references across utterances and scores confidence
- Core
- No dependencies
- Scripture store · verse lookup
- Web host · detected verses
Display pages
One styled page per projector or OBS output
- React
Scripture store
Bundled KJV, ASV and BSB for offline use
- SQLite
- Dapper
- api.bible · online translations
api.bible
Optional NIV, AMP and MSG, cached for 14 days
- HTTP
- Local cache
All connections
- Desktop launcher to Web host: starts, restarts
- Audio capture to Speech engine: speech chunks
- Speech engine to Reference parser: transcripts
- Reference parser to Scripture store: verse lookup
- Scripture store to api.bible: online translations
- Reference parser to Web host: detected verses
- Web host to Operator console: SSE + REST
- Web host to Display pages: SSE /events
Innovation highlights.
Innovation one
Context that bridges the pauses
Sticky book and chapter context survives a window of utterances, so a reference split by filler still resolves. An explicit cue can resume the last passage long after.
Innovation two
Confidence before the screen
Every heuristic match carries a confidence score for a confirm gate, because on a live display a false positive costs more than a miss.
Innovation three
Recognition biased to scripture
Whisper is primed with a vocabulary built from the parser's own book catalogue, so book names and numbers survive transcription.
Technology stack
What it is built with.
Host
- .NET 10
- ASP.NET Core Minimal API
- Server-Sent Events
Speech and audio
- Whisper.net
- sherpa-onnx
- PortAudio
Data
- SQLite
- Dapper
- api.bible
Interface
- React
- Vite
- Avalonia
Quality
- xUnit
- Golden-audio tests
Next project