AI as a detector, not a decision-maker
Designing Scribe AI’s live path: offline speech recognition, rule-based scripture detection, and a human approval step that keeps a wrong guess off a church screen.
Scribe AI listens to a church service and tries to catch scripture references as they are spoken — “turn with me to John chapter three, verse sixteen” — so the verse can be put on screen without a volunteer scrambling to type it. It is a small system, but it sits in an unforgiving place: whatever it gets wrong is projected in front of the whole room.
Two constraints shaped nearly every design decision: it must keep working without the internet, and nothing it detects should reach the screen without a person agreeing.
Offline is a requirement, not a feature
In the settings Scribe is designed for, a dropped connection in the middle of a service is ordinary, not an edge case. So everything in the live path runs as a local process against local data: microphone capture, speech recognition, detection, Bible lookup, and display. There is no network call anywhere in it.
Speech recognition uses Faster-Whisper’s small model on the CPU, quantised to int8, with greedy decoding (beam_size=1) and voice-activity filtering to skip silence. Audio is captured at 16 kHz and transcribed in five-second chunks. These are deliberate trade-offs for the kind of laptop a church production desk actually has: a smaller, faster model that keeps up with speech, rather than a more accurate one that falls behind it.
Rules, where rules are enough
It would have been easy to reach for a learned model to find references in the transcript. I did not, because the problem does not need one. Spoken references come in a small number of shapes, and the detector covers four:
John 3:16
John chapter 3 verse 16
John 3 16
John 316 ← how speech recognition often writes itThe patterns match against all 66 book names and common abbreviations. Every match is then checked against the real number of chapters in that book, so “Genesis 60” is rejected because Genesis has 50. That check matters most for the compressed form, where the transcript gives a run of digits and the detector has to decide where the chapter ends and the verse begins.
The benefit is not accuracy for its own sake. It is that every decision is inspectable: when the detector fires or fails to fire, you can read exactly why. There is no confidence score, and I have not pretended there is one. Validity is binary.
References don’t respect chunk boundaries
A preacher does not pause neatly every five seconds. “John chapter three” can land at the end of one chunk and “verse sixteen” at the start of the next. When a chunk produces no detection, the last 60 characters of its transcript are carried forward and prepended to the next one. When a chunk does produce a detection, the carried text is cleared, so the same reference is not detected twice.
The approval step
A detection never goes straight to the screen. It goes to an operator console, a separate always-on-top window where a person can approve it, dismiss it, or type a reference manually. Only an operator’s action — approving a detection or entering a reference themselves — puts a verse on the display.
This is the system’s answer to “what happens when the AI is wrong?” A misheard reference costs the operator one click. Without the gate, it costs a wrong verse in front of a congregation. The manual entry path also means the operator is never dependent on the detector: if it misses something, they can still put it up.
There is a smaller engineering detail behind that. Audio capture and transcription run on a background thread, while the console and display belong to the Tkinter UI thread. The audio thread never touches the interface directly; it schedules work on the UI thread with root.after, which keeps the two from fighting over the same widgets.
Being clear about what is not connected
Scribe also contains a service-state machine — keyword rules that track whether the service is in worship, sermon, prayer, or announcements — and a lyric detector that fuzzy-matches transcript text against a song library. Both are implemented and tested. Neither is connected to the live microphone path yet. They run together in a separate pipeline that is driven by simulated text.
I think saying that plainly matters more than it might seem. It would be easy to describe the module list and let a reader assume it all runs live. The lyric matcher’s threshold has only been tested on clean text, and the song library is placeholder data. Those are real limitations, and they belong in the description of the system, not in a footnote.
What comes next
- An evaluation set built from recordings of real services, with accents, crosstalk, and music, so detection is measured on the conditions it will face.
- One live entry point that includes state tracking and lyric detection, replacing the split between the live path and the simulated pipeline.
- An interface around speech recognition so other models — including ones better suited to African languages — can be swapped in and compared fairly.
The general lesson I keep returning to: the most important design decision in an applied AI system is often not the model. It is where the model’s output is allowed to go, and who gets to say no.
Originally published on DEV Community.