DocsAI Listener

AI Listener

The AI Listener is Spectra's headline feature — it detects Bible verse references from live speech and makes them instantly available for presentation. Detected verses go straight to the live screen, bypassing the slide deck and the schedule queue — the AI listener is its own presentation pipeline.

How It Works

Spectra uses Moonshine ASR, a high-performance speech recognition model that runs entirely on your local machine. Here's the pipeline:

1

Audio Capture

Your microphone captures audio in real-time. Audio is processed in small chunks for low latency.

2

Local Transcription

The Moonshine ASR model transcribes speech to text on your device. No audio ever leaves your machine.

3

Reference Detection

Spectra scans the transcription for Bible references using pattern matching that handles many spoken formats.

4

Verse Lookup

Detected references are matched to actual verses from the Bible database and displayed in the AI Listener panel.

5

Direct-Live Presentation

Click a detected verse (or let Autopilot do it) and the verse goes straight to the live screen — no slide creation, no schedule entry, no preview panel.

Privacy

Audio is processed in real-time and immediately discarded after transcription. Nothing is ever recorded, stored to disk, or transmitted to any server. Your sermons remain completely private.

Supported Reference Formats

Spectra recognizes a wide variety of Bible reference patterns from natural speech:

Spoken FormatDetected AsNotes
John 3:16John 3:16Standard format
First Corinthians chapter 131 Corinthians 13:1Spoken ordinals + “chapter” keyword
Genesis 1Genesis 1:1Chapter-only defaults to verse 1
Psalm 23 verses 1 through 6Psalm 23:1-6Verse ranges
Romans chapter 8 verse 28Romans 8:28Verbose spoken format

Chapter keyword locking

When the speaker says “chapter” explicitly (e.g., “chapter 2”), Spectra locks in verse 1 immediately. This prevents stray numbers in subsequent speech from being incorrectly interpreted as verse numbers.

Autopilot Mode

Autopilot turns the AI Listener into a fully hands-free operator. Flip the toggle in the Detected Verses header and every verse Spectra detects from your speech goes straight to the live screen — no clicks, no schedule, no manual intervention.

Spectra AI can make mistakes

Enabling Autopilot shows a disclaimer dialog. The AI may occasionally misrecognize a reference or surface the wrong verse. Review everything on screen and be ready to flip Autopilot off if needed — turning it off is instant and requires no confirmation.

Correction window

Autopilot holds each detection for a short quiet window (~800 ms) before pushing it live. If a second detection comes in during that window — for example, Spectra first hears “John 3” and ~200 ms later corrects to “John 3:16” — only the most recent one is presented. Wrong early guesses are overwritten by the correction.

Voice commands

While Autopilot is on, you can step through verses in the current chapter just by saying a command. No remote, no shortcut, no interruption to preaching flow.

CommandActionNotes
“next verse” / “next”Advance to the next verse in the same chapterBare “next” must be at the end of an utterance
“previous verse” / “previous”Step back one verseBare “previous” must be at the end of an utterance
“back” / “go back”Step back one verseSynonyms for “previous”

You don’t have to remember the exact words. Beyond the phrases above, Spectra understands the meaningof a short command, so natural variations like “advance”, “move forward”, “go to the next one”, or “take it back” step the verse too. This runs entirely on-device and only kicks in for short, command-like phrases — never on a sentence of preaching.

No false triggers

Compound phrases like “next time”, “next week”, “previous chapter”, or “go back to Genesis” will never trigger a step command. Exact keywords must be at the end of an utterance (or paired with “verse”), and the meaning-based matcher only considers short phrases — a full sentence of preaching that happens to contain “next” won’t move the slide.

Re-presenting a previous verse

If you've stepped through several verses with “next” and then quote an earlier one explicitly (e.g., back to John 1:10), Autopilot recognizes the re-detection and sends it back to the live screen immediately — even though it was already shown.

Out-of-range verse warnings

If you or the speaker quotes a verse that doesn't exist — for example, “John 2:28” when John 2 only has 25 verses — Spectra surfaces a toast like “John chapter 2 verse 28 does not exist. Has 25 verses.” so you know immediately without having to go hunt for why nothing appeared.

Full transcript side panel

The floating AI Listener widget has a transcript toggle (file icon) in the title bar. Click it to open a companion panel docked next to the main widget at the same height, showing the entire accumulated transcript for the session.

  • The panel auto-docks to the left of the main widget (or to the right if there's no room on the left).
  • Content wraps naturally and scrolls independently of the main widget.
  • A word counter in the header tracks the transcript length.
  • Future releases will add copy, export, and AI-refine actions directly from this panel.

Tips for Best Accuracy

Microphone Placement

  • Position the mic within 2-3 feet of the speaker for optimal clarity
  • A lapel/lavalier mic gives the most consistent results
  • If using a podium mic, ensure it's pointed toward the speaker's mouth
  • An external USB microphone significantly outperforms built-in laptop mics

Environment

  • Minimize background noise during detection — instruments, HVAC, and crowd noise reduce accuracy
  • If possible, use the mic before the room's speaker system to avoid echo
  • Smaller, quieter rooms naturally produce better results

Speaking Style

  • Clear enunciation of book names significantly improves detection
  • A brief pause before and after quoting a reference helps
  • Standard reference phrasing (e.g., “John chapter 3 verse 16”) is detected most reliably

Troubleshooting

“AI not ready”

This appears when the AI models are still loading. On first launch, Spectra downloads its models (~480 MB total — Moonshine ASR plus embeddings). After that, they take a few seconds to load into memory on each app start. Wait up to 2 minutes on first launch.

Warning

If the model fails to load after 2 minutes, try restarting Spectra. The app will automatically attempt to reload the model. If the issue persists, the model files may be corrupted — delete and re-download from the settings.

No verses being detected

  • Check that the correct microphone is selected in your system settings
  • Verify that audio levels are visible in the AI Listener panel (the waveform should move when there's speech)
  • On macOS, ensure Spectra has microphone permission in System Settings → Privacy & Security → Microphone
  • Try speaking a clear, common reference like “John 3:16” as a test

False or incorrect detections

Occasionally, the AI may misidentify spoken words as Bible references. If this happens frequently:

  • Ensure the microphone isn't picking up music or other non-speech audio
  • Background conversations can sometimes trigger false detections
  • Detected verses can be dismissed from the panel if they're incorrect