NoteScribe is a focused macOS voice transcription app for automatic speech recognition. It runs fully offline using Frontier Core ML audio models from Fluid Inference (Nvidia Parakeet TDT v2 and v3) and is designed for fast dictation into any text field or for transcribing audio/video files.
There are two release builds:
- Parakeet V3 (Multilingual - 25 European languages)
- Parakeet V2+V3 Combined (both models bundled; larger download)
V3 is the default release. If you want in-app model switching between V2 and V3 without separate installs, choose the combined build.
- Version:
v1.3.5(March 1, 2026) - Highlights: migrated the app to its new bundle identifier with forward migration for existing local data, and changed the uninstall helper to preserve user-created data by default.
- Release notes:
_docs/RELEASE_NOTES_v1.3.5.md - Downloads: GitHub Releases
Both models are Core ML models optimized for Apple silicon. In practice, Parakeet V2/V3 Core ML runs among the fastest offline ASR options on Mac today. Compared to Whisper GGML (base / large-v3 / turbo), these models show minimal extra RAM overhead, similar CPU thread usage, and a clear shift of work to the Neural Engine, leaving the rest of the system more responsive.
Compared to GGML models, the CoreML models use:
- ~50% less RAM,
- ~50% less energy draw,
- ~30-40% less CPU load (at peak)
- Hotkey dictation that pastes into the frontmost app
- File transcription (audio/video)
- History and replacements
- Download the DMG for v3 or v2v3 (combined) from GitHub Releases.
- Open the DMG and drag
NoteScribe.appinto Applications. - Launch the app and grant microphone (and accessibility for paste) when prompted.
An uninstall helper script is included in this repo: uninstall_notescribe.sh.
- macOS 14+
- Apple silicon (M-series). Intel Macs are not supported.
- Build/sign/notarize commands:
_docs/DEV_BUILD_COMMANDS.md - Current implementation + release packaging status:
_docs/CURRENT_STATE.md