AnchorCast v8.0.0 — Sermon Audio Studio & Scripture Detection Improvements

🎙️ MP3 Recording + ✂️ Audio Trimming + 🎧 Intro & Closing Voiceovers + 📖 Smarter Scripture Detection

AnchorCast v8.0.0 brings together a complete sermon-audio workflow and the Scripture detection, confidence-setting and Manual/Auto improvements prepared for the unpublished v1.7.2 draft. These changes are presented as one release: v8.0.0, not as a separate v1.7.2 update.

Record a sermon, trim unwanted audio from its beginning or end, and optionally add reusable introductions and closing messages with background music. The original recording stays intact. During the service, grouped Scripture cards, improved reference parsing and consistent confidence benchmarks make detection easier to manage without losing legitimate repeated quotations.

---

🎙️ New: Local Sermon Audio Recorder

Use Record Audio below Start Transcript to record the microphone selected in Settings → Audio & AI. Opening the controls does not start recording automatically. Audio recording and AI transcription are independent: starting, pausing or stopping one does not control the other.

Control	Behavior
Start Recording	Starts local microphone recording with an input-level meter and elapsed time.
Pause / Resume	Pauses or resumes capture; paused time is excluded from the audio.
Stop & Save	Finalizes the original recording before opening the optional finishing editor.
Save MP3 As…	Copies a saved recording to a chosen location and filename.
Open Recordings	Opens the local folder containing the recordings.

Audio is streamed locally into a genuine 128 kbps mono MP3, rather than keeping an entire sermon in memory. The recorder itself does not require registration, a cloud API or a Whisper model. AI Transcription retains its existing registration requirements. Interrupted files are retained where possible; listen to them before sharing.

Clear recording status

The main Audio button displays a pulsing red dot and REC HH:MM:SS, including when recorder controls are minimized. The recorder monitor has a matching indicator. Paused sessions show a steady amber pause marker. Startup, saving, stopped and error states do not falsely indicate active capture. Reduced-motion preferences keep the indicator steady; timer text remains readable without flashing.

Separate Minimize and Close

Minimize (−) hides only the recorder controls and leaves recording running; a paused session remains paused. Click the main Audio button to restore the same session.

Close (×) asks to Stop & Save, then Close when recording or paused. Cancel leaves the recording unchanged. Saving must complete successfully before the controls close; errors remain visible. Escape follows the close-confirmation path, and backdrop clicks do not stop or hide the recorder. Closing idle or already-saved controls needs no confirmation.

---

✂️ New: Trim and Finish Recording

Stop & Save saves the original MP3 first, then opens Trim and finish recording. For an earlier saved sermon, select it in the recorder and choose Trim / Add Intro / Closing….

Trim control	Purpose
Start at / End at	Specify the portion of the sermon to keep.
Start/end sliders	Adjust the beginning and ending visually.
Set start/end to playhead	Use the original audio player's current position.
Play kept section	Audition the selection, stopping at the chosen end.
Reset trim	Restore the complete original recording.

Times accept seconds, MM:SS or HH:MM:SS.mmm. Invalid, empty, reversed or out-of-range selections are blocked. The editor reports the original and retained duration and checks that the source has not changed since preview.

Only the sermon input is trimmed. Prepared narration and its background music are not shortened. This is a beginning/end trim editor, not a multitrack editor or a tool for removing sections from the middle of a sermon.

Optional introduction and closing

Both inclusion boxes start unchecked. Choose the result to create:

Selection	Finished result
Keep original only	Keep the already-saved original without creating a finished copy.
Trim only; no voiceovers selected	Export the selected sermon section by itself.
Add introduction	Introduction → selected sermon section.
Add closing message	Selected sermon section → closing.
Both	Introduction → selected sermon section → closing.

Create finished MP3 writes a separate 128 kbps mono MP3 under Recordings/Exports. Preview the result, use Save finished MP3 As…, or choose Open finished messages. The original and previous finished copies are not overwritten. To revise a finished message, return to its original recording and create a new copy.

Cancel remains available

A visible Cancel button stays in the footer, including when the dialog scrolls. Before export it closes the optional editor; during loading/export it cancels that dialog's operation, waits for cleanup and closes. Escape follows the same path. Cancelling after a finished file has already been saved does not delete it. Unrelated recording or voice-generation operations are not cancelled.

Export failures leave the editor open with an error and retain the trim selection for retry. Native application shutdown and interrupted-capture paths preserve original-audio handling rather than opening an optional finishing dialog during exit.

---

🎧 New: Reusable Introduction and Closing Voiceovers

Prepare clips in Settings → Audio & AI → Recording, then reuse them for later sermons.

• Editable starter scripts: generic introduction and closing examples without a church name or mandatory contact details. Use sample text asks before replacing existing nonempty text. Optional phone, email, platform and handle fields can fill placeholders in a customized script; unresolved placeholders block preparation.
• Formal male narration preset: Onyx is the default, with calm, formal voice direction, adjustable speed and other built-in voice choices. Preview the church name, contact details, pacing and pronunciation before publishing.
• Optional OpenAI generation: Generate voice uses the saved OpenAI API key and gpt-4o-mini-tts. Explicit consent is required; generation needs an internet connection and may incur API charges. Only the selected script, resolved contacts and voice direction are sent. This feature does not upload the preacher's recording.
• Import narration: use authorized audio recorded elsewhere instead of generating speech. Imported narration and subsequent mixing work locally without an OpenAI key. This implementation does not provide Claude-key speech generation; recognizable Anthropic keys entered in the OpenAI field are blocked locally before a request.
• Background music: includes Warm Welcome, a music-off option and authorized custom-track import. Music level, lead-in, ending and transition gap are adjustable, with fades around the narration. Music accompanies the introduction and closing only—not the sermon.
• Offline reuse: prepared clips are saved locally. Changing background music requires Rebuild mix; changing the script, relevant contact details or synthetic voice settings requires preparing the affected narration again. Outdated clips are marked rather than silently exported.

Your script only—no added spoken preface

Generated introductions and closings begin with the text you enter. AnchorCast does not prepend “This introduction is AI-narrated” or its closing equivalent.

Written listener-disclosure guidance appears in Settings and the finishing dialog for selected AI clips. It is not spoken or mixed into the MP3. Include an appropriate notice in the episode description or a visible listener-facing player notice when sharing generated narration. Imported-only clips and the original sermon are not falsely labelled as generated audio.

For users who tested earlier development builds, old generated clips containing the preface must be regenerated with Generate voice. Rebuild mix cannot remove speech already in a voice track. Those files are retained and marked for regeneration. A one-time settings update replaces only untouched former samples, with a local backup; customized scripts, contacts, imported audio and earlier finished messages are preserved.

Existing assets/audio folder

The bundled instrumental and its notice live in the project-root assets/audio folder; existing assets/Audio casing is supported. Generated/imported narration, custom tracks, recordings and finished exports remain in their writable user-data locations. The bundled-folder change alone does not require regenerating voiceovers.

---

🎯 Consistent 60%, 75% and 85% Confidence Benchmarks

The saved Settings → Audio & AI → AI Confidence Threshold controls both admission to the detection panel and Auto eligibility. The hidden 80% online floor, 88% Auto gate and conflicting legacy per-type thresholds no longer override the selected benchmark.

Setting	Benchmark
Low (60%)	More potential matches qualify.
Medium (75%)	Balanced default.
High (85%)	More selective; excludes candidates below 85%.

The minimum is inclusive: a score equal to the selected threshold qualifies for the confidence check. Existing valid selections are retained, and raising the threshold does not delete old History records. Scores are not averaged or artificially increased to make a group pass.

Confirmation remains separate from confidence. An incomplete reference or generic thematic suggestion does not become eligible for Auto simply because a model assigns it 95%. A current, confirmed 85% match can qualify when the selected benchmark is 75%; it is no longer blocked by the former hidden 88% rule. Confidence scores are heuristic match scores, not guarantees of real-world accuracy.

---

📚 Cleaner Scripture Cards Without Losing Repetitions

One ongoing Scripture discussion appears as one active card. An announced reference, its reading and continued explanation are grouped instead of producing several independent-looking cards, including the reported 1 John 1:9 and Philippians 1:6 examples.

Individual accepted records remain available in Review, History and Debug log, including source excerpts and timestamps. They no longer occupy the main canvas as recorded-match counts or expandable evidence rows. This is not a whole-service duplicate ban: a separately announced reference, a return after another confirmed passage or a longer gap can begin a new discussion.

Unrelated thematic matches appear in a separate, initially collapsed Suggestions / Review section. Supporting suggestions do not downgrade a confirmed card, replace its supported score or acquire Auto permission merely by being grouped with it.

Stable green indicators

A confirmed reference or quotation meeting the selected benchmark stays green, even after it becomes an earlier occurrence. Suggestions remain amber. Rejected, superseded or below-threshold evidence is not shown as confirmed green. Presentation age and readiness appear as text rather than being mistaken for lower confidence; the guard against automatically projecting stale evidence remains active.

Readiness and evidence updates no longer replay every card's entry animation or unnecessarily reset scroll position. Operator corrections, focused controls and expanded passages are preserved during redraws.

---

📖 Announced Scripture Ranges Stay Together

A passage such as Daniel 3:16–27 appears as one expandable card containing all 12 verses in ascending order, with Present verse 16, individual verse presentation and Schedule passage.

Auto initially queues the first verse, not all unread members. Long announcements no longer fill the eight pending slots and discard their opening verses. Subsequent explicit references or supported quotations can advance the reading; Auto does not race through an entire range just because it was announced.

Next verse follows the current reading position: after announcing 16–27 it advances from 16 to 17, not to 28. Only the opening verse is initially supplied to local transcription hints. Non-contiguous lists remain lists without invented intermediate verses, and the existing opt-in complete-passage display remains supported for contiguous ranges.

---

📖 More Reliable Reference Parsing and Passage Context

• No verse-1 guessing: “Hebrews chapter four” establishes chapter context without creating Hebrews 4:1. “Verse…” and incomplete number sequences remain pending before a verse card is created, in both Manual and Auto modes.
• Spoken digits and citation forms: “Psalm one one nine, verse one zero five” resolves as Psalm 119:105 rather than a premature Psalm 1:1. Ordinary “Romans one one” remains Romans 1:1. Singular “Hebrew,” chapter-before-book wording and book/number forms without spaces are supported.
• Chapter/verse corrections: “Psalm 89, I mean verse 89” can retain the earlier Psalm 119 context. Later corrections mark obsolete occurrences superseded rather than erase their history.
• Main reading versus side references: ambiguous bare verse numbers wait for explicit context or corroborating installed Bible text instead of inheriting the last side reference automatically. “Read on” alone does not invent a verse.
• Quotations without spoken references: distinctive quotation evidence can identify a Scripture. Generic thematic associations remain suggestions; ambiguous shared quotations are held outside the active reading.
• Stronger evidence: later words can improve the same occurrence and trigger another Auto eligibility check without creating an extra card. Rejected/dismissed cards and cleared panels are not silently reactivated.

---

🖥️ Manual and Auto Behave Consistently

Selecting Manual in Options now disables Auto. Toolbar changes are saved, and saving unrelated settings from an already-open Options window does not overwrite a newer toolbar choice when its mode control was left unchanged.

Switching to Manual, manually changing output, rejecting a match or clearing the panel/transcript cancels pending Auto work. Late results from earlier speech cannot override a newer manual choice or an operator-corrected reference.

Automatic presentation uses a brief settling period and controlled spacing between output changes. It rechecks mode, Live state, threshold, status and freshness before committing. Turning Auto on can reconsider the newest fresh eligible card without replaying an old backlog. Supporting phrases in one ongoing verse discussion do not repeatedly send that same verse to the live display.

In Manual mode, detections remain available for review and presentation requires an operator action such as Present.

🌐 Online Detection Reliability and Diagnostics

Online response handling reads all provider text blocks and validates one complete JSON array and its items. Valid empty arrays and harmless surrounding commentary no longer look like network outages. Truncated, ambiguous or unsupported results are rejected without inventing missing references.

Diagnostics distinguish response decoding, schema/evidence problems, truncation and actual HTTP/network failures. Parsing problems use a shorter retry cadence rather than the previous minute-long network backoff, while local detection continues independently.

Debug log separates grouped-card counts from accepted occurrence counts and includes discussion IDs, passage membership, effective threshold, mode, confirmation decisions, evidence upgrades, queue holds/cancellations and projection IPC results. A successful IPC request does not prove a physical projector displayed it. Review private transcript excerpts before sharing a log publicly.

---

⏱️ Existing Features Retained

Negative overtime with continuing red blinking, Hide on Projection / Show on Projection while the timer keeps running, operator timer monitoring, selected-translation loading, independent detection History/recovery and existing Theme Designer behavior are retained.

Users can continue into the basic app without registration. AI Transcription, additional Bible translations, NDI and Remote Control retain their existing registration requirements. Audio recording is independent of those protected features.

📦 Installer Variants

Choose the installer matching the target operating system and processor when publishing/downloading this release.

Platform	Variant	Description
Windows	Full	Python and Whisper model bundled for local transcription.
Windows	Light	Python bundled; downloads the Whisper model on first use.
macOS Apple Silicon	Full	ARM64 build with the Whisper model bundled.
macOS Apple Silicon	Light	ARM64 build; downloads the model on first use.
macOS Intel	Full	x64 build with the Whisper model bundled.
macOS Intel	Light	x64 build; downloads the model on first use.

⬆️ Upgrade and Build Notes

Upgrading from v1.7.1 or an earlier release: there is no separate v1.7.2 release to install first. The previously unpublished changes are included in v8.0.0. Application identity and data paths are unchanged; saved transcripts and historical detection lists are not automatically rewritten. Back up the project and application data before upgrading.

Development-build users: keep your prepared narration and original recordings. Old AI clips containing the spoken preface require regeneration; imported narration is unchanged. Rebuild a finished message from its raw original to use new trims or voiceovers. Untouched sample scripts can migrate once with a backup; custom scripts remain intact.

Building from source: the combined release uses ffmpeg-static. Install dependencies and synchronize the lockfile on the appropriate build machine before using npm ci, and package an encoder matching each target platform/architecture. Use the latest source setup and verification instructions in V8_INSTALL_AND_TEST_NOTES.md and RECORDING_TRIM_NOTES.md. Native merge scripts preserve unrelated configuration, including local SMTP settings, and keep backups outside the project.

Keep the existing assets, Bible databases, Python/models and build resources. Retain third-party notices and complete platform/signing checks. Optional text-to-speech uses the existing OpenAI key storage; no new encryption of that storage is claimed.

The source packages and this documentation are not EXE/DMG installers. Rebuild and test each installer before distribution. Publishing these notes does not itself publish or validate downloadable assets.

Validation Scope

The latest implementation results recorded in the trimming notes report 581 passing Node tests and 173 Chromium checks, including 58 new trim/backend/preview/installer tests and 37 new trim/cancel interface checks. These totals are not added to earlier feature-test counts.

MP3 decoding, trim duration, clip order, mixing and unchanged original-file hashes were tested using synthetic audio and Linux FFmpeg. Speech-provider responses, microphone permissions, Electron interfaces and native file dialogs were simulated. Browser preview used blob-fed test audio; loopback byte-range serving was checked separately in Node.

Not tested in those implementation results: live OpenAI generation or actual voice/pronunciation, native Windows/macOS installers, platform-specific encoder downloads, physical microphones, multi-hour recording reliability, physical projectors, NDI and remote devices. Preview generated narration and test a short native recording, trim, export and Cancel before use in a service. No universal transcription or detection accuracy guarantee is made.

This release-note/changelog consolidation changes documentation and matching documentation hashes only; it is not a new runtime test or code update.

Technical references retained from the source notes

• OpenAI text-to-speech guide
• OpenAI speech API reference
• FFmpeg audio filters