Local Meeting Transcript Mode

[asmit203]

So the feature in my mind is mostly is for transcribing meetings/ audio through the model

Add Local Meeting Transcript Mode

Summary

• Add a separate Meeting Capture mode that records system audio locally, transcribes it with the existing Whisper pipeline, and
  writes timestamped transcript logs into a user-selected folder.
• Keep current microphone dictation unchanged.
• In meeting mode, do not paste text into the focused app; only save transcript output.
• Store one transcript file per meeting session, growing incrementally while capture is running.

Key Changes

• Config and menu
  • Add meetingTranscriptDirectory: String? to Config.
  • Add tray actions: Start Meeting Capture, Stop Meeting Capture, Choose Transcript Folder…, Open Transcript Folder, Open
    Current Transcript.
  • Persist the chosen folder path in ~/.config/open-wispr/config.json.
  • Disable Start Meeting Capture until a writable transcript folder is configured.
• Permissions and packaging
  • Add ScreenCaptureKit and CoreMedia to Package.swift.
  • Add NSAudioCaptureUsageDescription to the generated app Info.plist.
  • Add screen-recording permission checks to Permissions using CGPreflightScreenCaptureAccess / CGRequestScreenCaptureAccess.
  • If permission is newly granted, require an app restart before first system-audio capture. This is the default behavior
    assumed by Apple’s sample.
• Capture pipeline
  • Introduce a capture abstraction so the current mic path and the new system-audio path share the same transcription pipeline.
  • Keep AudioRecorder for mic dictation.
  • Add a new SystemAudioCaptureSession that uses ScreenCaptureKit, captures audio only, excludes OpenWispr’s own audio, and
    ignores microphone input.
  • Capture audio continuously and cut it into fixed 30-second non-overlapping chunks.
  • Convert chunk audio to mono 16 kHz WAV before passing it to Transcriber.
  • Run chunk transcription on a single serial queue so appended transcript text stays in order.
  • Delete temporary chunk audio files after each transcription completes.
  • While meeting capture is active, ignore the dictation hotkey to avoid overlapping recorder/transcriber state.
• Transcript logging
  • Add a TranscriptLogStore responsible for folder validation, session file creation, and append operations.
  • Create one file per session named meeting-YYYY-MM-DD-HHmmss.md.
  • Write a small header at session start: start time, source System Audio, model, and language.
  • Append each successful chunk as one line: [HH:mm:ss] transcribed text.
  • Skip empty / blank-audio chunks after existing Whisper marker cleanup.
  • On stop, flush remaining queued transcriptions and append session end time.
• App orchestration
  • Extend AppDelegate with a separate meeting-capture state machine and start/stop handlers.
  • Extend StatusBarController.State with meeting-specific states so the menu and icon reflect “capturing meeting audio” vs
    normal dictation/transcribing.
  • Reuse existing Transcriber and text post-processing; bypass TextInserter for meeting mode.

Test Plan

• Add unit tests for Config decoding/saving of meetingTranscriptDirectory.
• Add unit tests for TranscriptLogStore:
  • creates session filenames correctly
  • rejects missing/unwritable directories
  • appends timestamped chunks in order
  • writes session header and end marker
• Add unit tests for chunk pipeline helpers:
  • chunk rollover at 30 seconds
  • blank chunk suppression
  • temp file cleanup after transcription
• Add manual validation for the system-audio path:
  • first-time permission flow for Screen Recording / system audio
  • folder chooser persists across restart
  • transcript file grows during an active Zoom/Meet playback
  • no text is pasted during meeting mode
  • normal mic dictation still works after stopping meeting mode

Assumptions And Defaults

• V1 is a separate mode, not a source toggle inside the existing dictation flow.
• V1 is save-only; meeting text is never auto-pasted.
• V1 stores one transcript file per session.
• V1 does not retain raw meeting audio after chunk transcription; only transcript logs persist.
• The chosen folder is stored as a plain filesystem path, not a security-scoped bookmark, because this app is currently
  distributed as an unsandboxed macOS app.
• V1 captures system output broadly and does not add per-app or per-window capture selection yet.

References

• Apple ScreenCaptureKit overview: https://developer.apple.com/documentation/screencapturekit/
• Apple sample for capturing screen/audio content on macOS:
  https://developer.apple.com/documentation/ScreenCaptureKit/capturing-screen-content-in-macos
• Apple media-capture authorization docs, including NSAudioCaptureUsageDescription:
  https://developer.apple.com/documentation/bundleresources/requesting-authorization-for-media-capture-on-macos

[asmit203]

I am currently implementing this issue.

[asmit203]

https://github.com/asmit203/open-wispr

I have this implemented here but the issue is I am unable to test it on my local machine as only homebrew tap is working properly and when I tried to use the following command I couldnt properly use the 0.34 from the main branch and got thown into 0.9 version

 2424  brew uninstall --formula open-wispr 2>/dev/null || true
 2425  brew install --build-from-source ./Formula/open-wispr.rb
 2426  pgrep -alf open-wispr
 2427  brew tap-new asmit/local
 2428  mkdir -p "$(brew --repository)/Library/Taps/asmit/homebrew-local/Formula"
 2429  cp open-wispr/Formula/open-wispr.rb \\n    "$(brew --repository)/Library/Taps/asmit/homebrew-local/Formula/open-wispr.rb"
 2430  brew uninstall open-wispr 2>/dev/null || true
 2431  brew install --build-from-source asmit/local/open-wispr
 2432  brew services start open-wispr
 2433  brew services start asmit/local/open-wispr
 2434  brew services restart asmit/local/open-wispr
 2435  brew services stop asmit/local/open-wispr
 2436  brew install open-wispr
 2437  brew service restart open-wispr
 2438  brew services restart open-wispr

[human37]

Have you tried running the scripts/dev.sh command? Thanks!

[m13v]

nice design doc. we built a similar system audio capture pipeline with ScreenCaptureKit for session recording, and a few notes that might help whoever implements this:

the 30-second non-overlapping chunk approach works but you'll get clipped words at chunk boundaries. adding even 1 second of overlap between chunks and then deduplicating the overlapping transcription segments gives much cleaner results. whisper handles the overlap gracefully, you just need to trim the first ~1 second of each chunk's transcript.

for the mono 16kHz conversion, SCK delivers audio as 48kHz float32 interleaved by default. converting in the SCStreamOutput delegate callback (before writing to the temp file) avoids an extra disk read/write cycle. AVAudioConverter handles the format conversion efficiently.

on the permission flow, CGRequestScreenCaptureAccess() is a one-shot call. if the user dismisses the dialog without granting, you cannot re-prompt without the user manually going to System Settings. worth detecting this state and showing a deep link button to the Screen Recording preference pane.

the restart requirement after granting screen recording permission is real and catches a lot of developers off guard. caching a flag in UserDefaults when permission was not granted at launch, then checking again on app activate, gives you a clean way to detect when the user has granted it and prompt for restart.

[m13v]

our ScreenCaptureKit service that handles system audio capture, format conversion, and permission flow: https://github.com/m13v/macos-session-replay/blob/main/Sources/SessionReplay/ScreenCaptureService.swift

and the transcription pipeline with chunked audio processing and model management: https://github.com/m13v/fazm/blob/main/Desktop/Sources/TranscriptionService.swift

[asmit203]

Have you tried running the scripts/dev.sh command? Thanks!

Have tried worked for once then it completely broke I ill try tomorrow again [Busy Today]

will also look into #48 (comment) & #48 (comment)

[m13v]

no worries, take your time. if it breaks again after running dev.sh once, check if there's a stale process still holding the audio device. sometimes the capture session doesn't clean up properly on crash. lmk how it goes

[asmit203]

no worries, take your time. if it breaks again after running dev.sh once, check if there's a stale process still holding the audio device. sometimes the capture session doesn't clean up properly on crash. lmk how it goes

It worked I did update the permission workflow and permission clearling while uninstall script runs now it does ask more properly to grant permission and it works in my machine and I did draft the pull request but I want the previous pull req at #42 to get merge first I guess

I dont know it will be nice you try that draft pull request out and lemme know if it is working properly simply play a youtube music video to check if it is working or not

[m13v]

nice, glad it's working. the permission cleanup on uninstall is a good call, most macOS audio apps miss that. makes sense to wait for #42 to land first. lmk if you need any eyes on the PR once it's ready for review.

[asmit203]

nice, glad it's working. the permission cleanup on uninstall is a good call, most macOS audio apps miss that. makes sense to wait for #42 to land first. lmk if you need any eyes on the PR once it's ready for review.

There were some conflicts in the #42 so I took it and added all the features into my dev branch of my fork, basically copy over those changes and that one is also working so i am waiting for Ammon to say something if he is okay with the copy over then I will make another PR and delete the #50 for a faster way.

[ Sorry about the re implementation of same thing on my fork and branch basically I needed the dictionary badly so I copied it over so I understand if you dont want it I can copy over the remaining changes for the #46 as I already implemented that too on my dev branch after it is merged (I got time today to completed all of it...)]

[m13v]

nice, sounds like you got it all consolidated. using codex to port changes across forks is a solid move. hope Ammon gets back to you soon, the meeting transcript feature is going to be really useful once it lands.