Skip to content
ajbarryiiiPublic

About

LocalFlow: local-only macOS dictation with a bundled ternary speech model. Based on FreeFlow by Zach Latta and contributors.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

LocalFlow icon

LocalFlow

⬇ Download LocalFlow beta DMG
Apple Silicon · macOS 26 or later
Developer ID signed · Apple-notarized beta


A native macOS menu-bar dictation app using the bundled LocalFlow English speech model, derived from NVIDIA Parakeet v2. This fork runs speech recognition on the Mac, with no API key, provider account, network transcription, LLM cleanup, translation, Edit Mode, or screenshot/app context capture.

Requires an Apple Silicon Mac running macOS 26 or later. Other devices are not supported by this model bundle yet. The app and model are about 388 MB together; runtime memory use can exceed the model's disk size.

Dictation

  1. Open the bundled app and complete local setup.
  2. Grant Microphone and Accessibility access. On macOS 27, the Accessibility permission pane is named Device Control and Data Access under System Settings → Privacy & Security. Screen Recording is not needed.
  3. Choose Hold to Talk, Tap to Toggle, and optional Paste Again shortcuts.
  4. Dictate into any text field. LocalFlow transcribes locally and pastes at the cursor, optionally restoring the previous clipboard.

The model prepares in the background at startup and stays in memory for the session. First device preparation can take several minutes; Settings and the menu show its state. Settings provides a retry if preparation fails. Recordings use independent chunks of up to 30 seconds once optimization completes, or 15 seconds while it is still in progress, and can lose context at boundaries. There is no cloud fallback.

Voice macros match complete phrases locally and paste predefined text. An optional trailing “press enter” command submits after pasting. Neither feature uses an LLM. Dictated instructions otherwise remain literal text.

The latest 20 recordings and transcripts are retained locally for playback and local retry. Clear or delete them from Run Log. Existing history and shortcut preferences remain compatible; old provider preferences and credentials are ignored and are not read or automatically erased.

Build locally

LocalFlow uses swiftc and Make, without Swift Package Manager or an Xcode project. Install Xcode and its command-line tools. Model weights are kept outside Git; obtain or prepare the pinned model bundle as described in the integration notes.

SDKROOT="$(xcrun --sdk macosx --show-sdk-path)" make check
git diff --check
SDKROOT="$(xcrun --sdk macosx --show-sdk-path)" make \
  ARCH="$(uname -m)" CODESIGN_IDENTITY=- PARAKEET_BUNDLE_DIR="MODEL_BUNDLE"
open "build/LocalFlow Dev.app"

The ad hoc development build may need its existing system approval refreshed when its executable changes. The app's permission guidance can reveal the exact running bundle in Finder for this repair.

LocalFlow uses its own bundle identifiers (com.ajbarryiii.localflow, and com.ajbarryiii.localflow.dev for development builds), so it never shares preferences, permissions, history, or audio with an installed copy of FreeFlow. Its data lives in ~/Library/Application Support/LocalFlow (or LocalFlow Dev), and external tools can poll is-recording in that directory.

The software updater checks this fork and may contact GitHub when checking or downloading app updates. Audio and transcripts are not sent by the updater. Turn off automatic update checks in Settings if desired.

Credits and licensing

LocalFlow is based on FreeFlow by Zach Latta and contributors. The original copyright and MIT license are preserved, with attribution included in Settings and every built app bundle. Application code retains the MIT license. Model and adapted inference-code licenses and attribution are included in Resources/Parakeet.

About

LocalFlow: local-only macOS dictation with a bundled ternary speech model. Based on FreeFlow by Zach Latta and contributors.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages