It listens to a sermon, lesson, or Bible discussion and puts every passage the speaker mentions on screen, in real time. Four detection lanes work together: spoken references ("John 3:16", "Romans, the eighth chapter"), passage following (announce a chapter and just read — it walks verse by verse and follows jumps), verbatim search (a quoted verse with no reference is found by its words across all 31,102 verses), and an AI lane that resolves half-remembered quotes — every AI answer is verified against the real verse text before anything is shown. If a detection can't be verified, nothing is shown.
Press Play the guided demo — a three-minute narrated tour where every feature fires live on screen. Then press Go live with my microphone and preach, teach, or read.
The main page is the operator's view: live transcript, detection feed, controls. The Audience screen button switches to a clean full-screen view for the projector — just the verses. Controls hide when the mouse is still and return when it moves; Escape returns to the operator view. For streaming software (OBS), add the page with ?obs=1 as a browser source for a transparent verse overlay.
As the preacher reads a displayed verse aloud, the engine compares the spoken words word-by-word against every licensed translation. When three verses each match one translation nearly word-for-word, the display switches to that translation and stays. Memory quotes and paraphrase never trigger a switch, and translations outside the licensed set (NKJV, CSB, The Message…) can't be matched — a note tells you the closest available translation is being shown instead.
Yes. With an empty .env the app still runs: the bundled public-domain King James Version provides the Bible text, and speech recognition falls back to your browser's built-in recognizer (Chrome and Edge support it best). The spoken-reference, passage-following, and verbatim-search lanes all work in this mode. Adding keys upgrades it: a YouVersion Platform key unlocks the licensed translation catalog, an AssemblyAI key unlocks the high-accuracy microphone, Tyndale/Crossway keys add NLT and ESV, and Gloo AI credentials enable the paraphrase lane.
Everything your YouVersion Platform key is licensed for (NIV, NASB, Amplified, and many more), plus the bundled KJV, plus NLT and ESV through their publishers' APIs. Spanish sermon mode swaps the pickers so a Spanish translation leads and English becomes the optional second line — and the bilingual toggle shows both at once.
Say the reference out loud. A spoken citation ("John 3:16", "First Corinthians 13", "Romans, the eighth chapter, verse 28") is the fastest lane in the engine — the verse is typically on screen within a second or two, often before you finish the sentence. Preachers who know this lead the screen naturally: announce the passage the way you already do — "turn with me to Psalm 23, verse 1" — and it's up before your congregation finds the page. Spoken ranges work too: "verses 1 through 3" puts the whole span up at once (a longer range opens with its first verses, and the screen follows as you read on). Announcing just a chapter puts up a title card immediately while people turn there; the moment you start reading, the screen follows verse by verse, including jumps ("down to verse 14"). You never need to wait for the engine to figure out where you are — but if you don't announce anything, the other lanes will still find you.
By design, the engine prefers silence over guessing. Fragments that can't be verified against real verse text, references the recognizer garbled, and AI resolutions that fail verification are all discarded. The detection feed on the operator view shows everything that was caught, with how and how fast.
The engine acts on live, in-progress transcription, which the recognizer sometimes revises mid-sentence. When the words themselves make the true passage clear (say the preacher cites First Peter but reads Hebrews), the engine corrects the screen to what was actually read.
Each session writes a local log file (transcript events, detections, timing) under logs/ on the machine running the server. They exist so operators can review what was said versus what was shown, and they never leave that machine. Delete the folder anytime, or disable logging by removing write access — the app runs fine without it.
The software is free (MIT). The optional third-party APIs have their own free tiers and pricing; the app includes daily budget caps and per-route rate limits so a public deployment can't run away with your quota.
The translation service is momentarily throttled or unreachable and the app is running on its offline floor. It recovers automatically; anything the server has fetched before is cached durably and keeps working during outages.
Yes — that's why it's MIT-licensed. Read the Terms for your responsibilities (especially third-party API terms and audio consent), the Privacy Policy for what data goes where, and the repository README for setup: github.com/dacini/red-letter-live.
Red Letter Live is open source (MIT), created by David Copercini · github.com/dacini/red-letter-live