Bible verse detection, while the sermon is still happening
NaveLight listens to the preaching and finds the passage before your operator can type it. A spoken reference resolves in about two seconds. When the preacher describes a passage without citing it, it finds that too. Every suggestion arrives with a confidence figure and a note of how it was found, and it all runs on your own computer.
Let me read it to you the way Paul wrote it.
Go with me to Romans chapter eight, verse twenty-eight
DETECTED IN 1.8s · KEYWORD MATCHES NEVER AUTO-AIR
How it decides
Nine stages, in this order, every time
“AI finds the verse” is a claim anyone can make. This is what actually runs, in sequence, and where each stage's confidence is capped. Cheap and certain first; expensive and speculative last, with the speculative stages barred from the screen.
- 010.97 · 0.85 when fuzzy
Explicit reference
A deterministic parser runs before any model is consulted. “Romans chapter eight, verse twenty-eight” or “turn to John three sixteen” resolves here, handling spoken numerals, shortened book names and mid-sentence corrections.
- 02Moves within the live song only
Lyric follow
If a song is already live, the words being sung move the slide within it. This never brings up anything new — it only moves inside what you already put on screen.
- 03Holds the current passage
Reading lock
Once a passage is up and being read aloud, detection stays with it rather than jumping at every incidental phrase.
- 04match · best guess
Song recall
Sung words are matched against your library, so an unplanned song or a reprise gets offered. It tells you whether that is a confident match or a best guess.
- 05Suppresses, never fires
Live-only gates
During a service two filters apply that don't apply to typed commands: a post-air hold, so something you just took down isn't immediately put back, and a prayer-speech filter, so it stays quiet during the prayer.
- 06Learned from your dismissals
Learned phrases
Verses you have dismissed as wrong stop being offered, and phrasing your church uses gets recognised over time. This is local learning on your machine, not a model trained on your services.
- 07Capped at 0.95
Story index
A curated narrative index for the passages preachers describe rather than cite — the prodigal son, the storm on the lake, the walls at Jericho.
- 08Ranked, never certain
Semantic search
On-device embeddings across the Bible, for paraphrase and theme when nothing above has matched.
- 09Capped at 0.60 · never auto-airs
Keyword fallback
Last resort, and treated as such. A candidate must share at least three words with the phrase, including at least one distinctive word — one that appears in roughly 1% of verses or fewer, so “the Lord said” cannot trigger anything.
The ordering is the point. A deterministic parser is faster and more trustworthy than a model, so it goes first and the models only see what it couldn't answer. The stages most likely to be wrong are the ones least able to act.
Who decides
Detection and projection are different things
Finding a passage is a guess with a number attached. Putting it in front of a congregation is a decision. NaveLight keeps those separate: detection always runs, and what reaches the screen is set by you, one capability at a time.
It ships cautious — the stages that guess are on Assist until you say otherwise — and one button returns everything to suggest-only mid-service if a morning starts getting away from you.
Smart Auto projects confident references & commands (≥95%); semantic matches always wait for you. Lyric follow never airs new content, it only moves within the song already live.
Straight answers
How the detection works
What software automatically detects Bible verses during a sermon?
NaveLight detects Bible verses live from the spoken sermon. A spoken reference resolves in roughly two seconds at about 97% confidence, and when the preacher describes a passage without citing it, a narrative index and on-device semantic search find it anyway. Detection runs on the church's own computer, so it works with no internet connection.
How accurate is automatic scripture detection?
Accuracy depends on how the passage was referred to, which is why NaveLight publishes a confidence figure with every suggestion rather than a single accuracy claim. Explicitly spoken references resolve at about 97%; narrative matches are capped at 95%; keyword matches are capped at 60% and can never put themselves on screen. The operator always sees both the number and how the match was found.
Does it put the wrong verse on screen?
Not by itself, unless you have allowed it to. NaveLight ships with the guessier stages set to suggest-only, so they appear in a list for the operator rather than going to the output. Even on Smart Auto, only matches at 95% confidence or above are projected, and keyword matches are excluded entirely.
Does it work if the preacher doesn't say the reference?
Yes, that is the harder case and the one it is built for. Most preachers describe a passage rather than cite it, so NaveLight runs a curated narrative index and an on-device semantic search over the text to find what is being described. Those matches are labelled “Meaning” so the operator knows to glance at them rather than trust them blindly.
Is the sermon audio sent anywhere for processing?
No. Transcription and every detection stage run on the church's own machine. There is no cloud service, no account and no per-minute cost, which is also why detection keeps working when the building's internet doesn't.
Still have a question?
Ask us anything, including the awkward ones. A real person answers, and if NaveLight isn't right for your church we'll say so.
Ready to put every word on cue?
Get on the list for early access. Or bring your media team along and we'll walk you through it properly.
No card, no account, no follow-up sequence. One email, when it's ready.