openbezal/rhema

What it solves

Rhema is a real-time AI-powered system designed for live sermons and broadcasts. It solves the problem of manually searching for and displaying Bible verses during a live talk, by automatically detecting references and quotes in the audio feed and rendering them as professional, broadcast-ready overlays via NDI.

How it works

The application uses a multi-strategy detection pipeline to identify verses:

• Speech-to-Text: It transcribes live audio using either local Whisper models or the Deepgram cloud API.
• Detection Strategies: It employs direct reference parsing (Aho-Corasick and fuzzy matching), quotation matching against known texts, and semantic search using Qwen3-0.6B ONNX embeddings for paraphrase matching.
• Context Tracking: It tracks sermon context and utilizes a "reading mode" that locks to specific books and chapters for easier navigation.
• Output: Detected verses are sent to a broadcast queue and rendered as overlays with customizable themes via the NDI 6 SDK.
• Backend/Frontend: Built with a Rust backend (Tauri v2) for performance-critical tasks like audio capture and ML inference, and a React frontend for the user interface.

Who it’s for

• Live production teams for churches and ministries.
• Broadcasters who need automated, real-time scripture reference overlays.
• Speakers who want voice-controlled translation switching and automated verse display.

Highlights

• Real-time STT: Support for both local (Whisper) and cloud (Deepgram) transcription.
• Cros-platform: Works on Windows, macOS, and Linux.
• Semantic Search: Optional embedding-based re-ranking for high-accuracy paraphrase detection.
• Broadcast Integration: Professional NDI output with configurable resolution and alpha modes.
• Voice Control: Ability to switch Bible translations (e.g., "read in NIV") via voice commands.
• Extensive Database: Bundled with 10 translations and over 340,000 cross-references.