Lorre

Lorre is a macOS transcription workspace for capturing or importing audio, processing it locally, reviewing speaker-labeled transcript segments, dictating into other apps, and exporting the result in multiple formats.

Key features

• Capture any conversation your way with Microphone, System audio, or Microphone + system audio.
• See live transcription as you record, with selectable Parakeet EOU or Nemotron streaming previews for supported live workflows.
• Powered by selectable Parakeet TDT 0.6B v3 multilingual batch ASR and Parakeet TDT 0.6B v2 for English-only batch transcripts, running on Apple's Neural Engine right on your Mac.
• Supports multilingual Parakeet v3 transcription, with explicit batch language hints for English, French, German, Spanish, Italian, Portuguese, Dutch, and Polish.
• Built on FluidAudio for local speech recognition, voice activity detection, and speaker diarization.
• Keep your recordings and transcripts on your Mac for a private local workflow.
• Enable Privacy Mode to automatically delete source audio after the transcript is saved.
• Import existing audio files and run them through the same transcription pipeline.
• Review transcripts with speaker labels, playback, speaker reassignment, and inline text editing.
• Use Global Dictation from any app to record a short microphone snippet, transcribe it locally, and insert the result into the focused text field.
• Let Call Watcher detect likely calls in apps such as FaceTime, Zoom, Microsoft Teams, Slack, Discord, Webex, and supported browser calls. Lorre always asks before recording and can show the prompt as a macOS notification or inside the app.
• Automatically write a Markdown copy after processing, with a configurable destination and filename template based on the transcript, date, source, language, duration, and speaker count.
• Export finished sessions as Markdown, plain text, or JSON.

What is FluidAudio?

FluidAudio is the on-device speech engine behind Lorre. It is a Swift library for Apple devices that combines speech-to-text, voice activity detection, and speaker diarization in one local pipeline.

Lorre defaults to FluidAudio's Parakeet TDT 0.6B v3 model for final transcription, and English-only workflows can choose Parakeet TDT 0.6B v2 as a batch option. In simple terms, Parakeet is the ASR model that listens to your recording and turns spoken words into text. FluidAudio's v3 model supports 25 European languages; Lorre exposes explicit batch language hints for English, French, German, Spanish, Italian, Portuguese, Dutch, and Polish. Lorre can also use FluidAudio's faster Parakeet EOU or Nemotron streaming models to show a live preview while you are still recording.

ANE-optimized means FluidAudio is tuned to run efficiently on Apple's Neural Engine, the part of Apple silicon designed for AI workloads. For a user, that usually means faster transcription, lower power use, and less pressure on the CPU and GPU.

Parakeet 0.6B means the model has about 600 million parameters. That is relatively compact compared with many modern AI models, which helps keep it practical for local use on a Mac without needing the kind of memory larger cloud-style models often expect.

The main benefit of FluidAudio is that it gives Lorre a complete local speech stack:

• your audio can stay on your Mac instead of being sent to a cloud API
• it is built for Apple devices, so it can take advantage of Apple hardware for speed and efficiency
• it handles the hard parts together: detecting speech, transcribing it, and separating speakers

That is what lets Lorre offer private, on-device transcription with speaker labeling in a single app.

Architecture

At a high level, Lorre routes recording, import, review, and export workflows through a SwiftUI app model, then keeps processing and storage local to the Mac.

flowchart LR
  input["Audio input<br/>microphone, system audio, imports, dictation"]
  signals["Local macOS signals<br/>focused field, apps, windows, device use"]
  ui["SwiftUI app<br/>recorder, shelf, transcript editor"]
  workflow["AppViewModel<br/>session and workflow state"]
  services["Core services<br/>record, play, process, export"]
  engines["On-device engines<br/>FluidAudio + macOS media frameworks"]
  storage["Local storage<br/>sessions, audio, transcripts, speakers, settings"]
  exports["Exports<br/>Markdown, text, JSON, automatic Markdown"]

  input --> ui
  signals --> workflow
  ui --> workflow
  workflow --> services
  services --> engines
  services <--> storage
  storage --> ui
  services --> exports

Loading

All session data lives under ~/Library/Application Support/Lorre/ unless you export it.

Configuration

Lorre keeps app preferences in ~/Library/Application Support/Lorre/settings.json. Most settings affect new recordings, imported audio, or retry-processing runs; they do not rewrite an already finished transcript unless you process that session again.

Recorder options

Option	Values	Default	Notes
Capture mode	Microphone, System audio, Microphone + system audio	Microphone	System-audio modes require Screen and System Audio Recording permission. When system audio is included, Lorre opens the native macOS picker so you can choose the app, window, or display audio to capture.
Live transcript preview	On or off	On when supported	Shows streaming text while recording. The final transcript is still produced by the selected batch transcription mode after recording stops.
Live engine	Parakeet low latency, Parakeet balanced, Parakeet high accuracy, Nemotron balanced, Nemotron high accuracy	Parakeet balanced	Controls the streaming preview model and latency/quality balance. Live-engine changes are locked while a recording is active.
Retention	Keep Audio, Delete After Transcript	Keep Audio	Delete After Transcript is privacy mode: Lorre deletes the source audio after the transcript is saved and keeps transcript data and exports. Playback and waveform review require retained audio.

Automation and system integration

Option	Values	Default	Notes
Global Dictation	On or off	Off	Press the configured shortcut from another app to start listening. Press it again to stop, transcribe locally, and insert the result into the text field that was focused when dictation started. Password and secure-input fields are rejected.
Dictation shortcut	Option-Shift-D, Option-Shift-Space, Command-Option-Shift-D, plus legacy Control-based choices	Option-Shift-D	Requires Microphone and Accessibility permission. Dictated text remains available to copy when insertion fails.
Call Watcher	On or off	Off	Uses local running-app, visible-window-title, camera, and microphone-use signals while Lorre is open. It recognizes native communication apps, Teams PWAs, and supported browser call titles. Detection never starts recording automatically.
Call recording source	Microphone, System audio, Microphone + system audio	Microphone + system audio	Used only after you accept a call prompt. Dismissed prompts have a cooldown before Lorre asks again for the same call.
Automatic Markdown export	On or off, plus destination folder	Off	Writes a Markdown file after a transcript completes. A destination folder must be selected before the option can be enabled.
Automatic export filename	Template string	{date}-{smart_title}.md	Supports {date}, {time}, {datetime}, {session_title}, {keywords}, {smart_title}, {source}, {language}, {duration}, {speaker_count}, and {session_id_short}. Invalid filename characters are sanitized.

Speech and processing options

Option	Values	Default	Notes
ASR mode	Parakeet v3 multilingual, Parakeet v2 English, Parakeet + Cohere	Parakeet v3 multilingual	Parakeet v3 multilingual is the default timed transcript path. Parakeet v2 English is English-only. Parakeet + Cohere adds an alternate draft while timed transcript rows still come from Parakeet.
Batch language	EN, FR, DE, ES, IT, PT, NL, PL	EN	Changing language away from English automatically falls back from the English-only v2 path to the multilingual v3 path.
Speaker recognition	On or off	On	Enables diarization during processing so Lorre can assign speaker labels automatically. Turning it off keeps speakers manual unless you retry processing later.
Diarization engine	Offline VBx, Sortformer, LS-EEND	Offline VBx	Offline VBx favors full-recording offline clustering. Sortformer targets stable identity continuity for up to 4 speakers. LS-EEND is overlap-friendly and supports up to 10 speakers.
Expected speakers	Auto, Exact 1, Exact 2, Exact 3, Exact 4, Range 2-4, Range 2-6	Auto	Gives the diarizer a speaker-count hint for new processing runs.
Known speakers	Local speaker name plus an audio sample	None	Enrolled voices are stored in known-speakers.json and known-speaker-samples/, then used to relabel diarization clusters and provide live speaker hints when supported. You can add, update, or delete enrolled speakers.
Vocabulary boosting	On or off, plus custom terms	Off	Available only when the linked runtime supports it. Terms are stored even in builds where boosting is inactive. Use one term per line, or Canonical: alias1, alias2 for variants.
Show confidence	On or off	Off	Shows confidence percentages under transcript rows when the ASR result includes confidence data.
Diarization debug JSON	On or off	Off	Writes diarization-debug.json into each processed session folder for troubleshooting speaker-label issues.
Model registry	Default Hugging Face registry or a custom base URL	https://huggingface.co	Set this before preparing or downloading models if you need a mirror or private registry. Leave blank to use the default.

The Prepare Models action downloads and warms the currently selected local model set. It stores the last-prepared status in settings.json, but it is an action rather than a processing preference.

Advanced settings file note: batchTranscription.parallelChunkConcurrency defaults to 4 and is normalized to a range of 1 to 8. The current UI preserves this value but does not expose a control for it.

Workspace and review options

Option	Values	Where it is stored
Session folders	Create, rename, delete, move sessions to a folder or Unfiled	Folder definitions live in settings.json; each session's folder assignment lives in its session.json. Deleting a folder moves its sessions to Unfiled.
Sidebar expansion	Expanded/collapsed view filters and folders	settings.json
Session title	Rename any recording	The session's session.json
Session notes	Add, save, or clear a private note	The session's session.json
Transcript text	Edit transcript segments inline	The session's transcript.json
Speaker labels	Reassign a segment or rename a speaker in the transcript	The session's transcript.json
Playback rate	0.75x, 1.0x, 1.25x, 1.5x	Current app session only
Export format	Markdown, Plain Text, JSON	Exported files are written to the destination you choose and export history is recorded in session.json.

Requirements

• macOS 14 or later
• Microphone access for microphone recording
• Screen and System Audio Recording access for system-audio capture
• Accessibility access for inserting Global Dictation text into other apps
• Notification access for Call Watcher prompts outside Lorre; if notifications are unavailable, Lorre falls back to an in-app prompt
• A local build environment for Swift if you want to run from source

Build and run

swift run

To build the app bundle in dist/:

./scripts/package_macos_app.sh

For an optimized release bundle:

./scripts/package_macos_app.sh release

To run the test suite:

./scripts/swift_test.sh

To run the local CI check used by this repo:

./scripts/ci_check.sh

Privacy and local data

Lorre stores session data in ~/Library/Application Support/Lorre/.

Each session is kept in its own folder and can contain:

• session.json for session metadata
• transcript.json for transcript data
• recorded or imported audio
• optional microphone and system-audio stem files for mixed recordings
• exported transcript files

If privacy mode is enabled before a recording or import, Lorre deletes the source audio after transcription completes and keeps the transcript and exports.

Global Dictation captures microphone audio only after you invoke its shortcut. Call Watcher does not capture audio: while enabled and Lorre is open, it evaluates local app identifiers, visible window titles, and camera or microphone usage to decide whether to show a prompt. Neither feature sends those signals or transcript data to a cloud service.

Export formats

• Markdown
• Plain text
• JSON