Vellum for macOS · Early release

Speak. Keep your words on your Mac.

Vellum turns German speech into text directly in your focused text field. Dictation and optional correction run locally on your Mac. Free, without an account or subscription.

macOS 14 or later · Apple Silicon · German dictation

Text where you work
Hold ⌥Z and speak. Vellum inserts text during dictation into supported text fields. The shortcut can be changed in Settings.
Correction is your choice
Enable or disable local correction in Settings. With the optional model installed, ⌥⇧Z also corrects selected German text. Review the result: spelling and grammar corrections can be incomplete.
A small presence
A movable floating pill shows recording and processing status. Double-click it to open Vellum. Closing the main window keeps the app available in the menu bar.
Control over your history
Transcripts are saved locally by default. Disable transcript storage in Settings if you do not want a history. Recordings and dictated text are not uploaded for recognition or correction.

Offline after the model download.

Download the speech model once in Vellum. Recognition and, with the separate correction model installed, correction run on your device, including without an internet connection. Model downloads use Hugging Face; Vellum checks the downloaded files against their SHA-256 checksums.

For work that needs discretion.

Notes for doctors and lawyers, interview drafts for journalists, research notes and everyday writing. Local processing helps keep spoken content on your Mac. Review names, numbers and specialist terms before using or sharing a transcript.

One package. Two local components.

Vellum Deutsch Lokal is the public PJ Labs ONNX package. Speech recognition is required; text correction is optional and can be enabled or disabled separately.

Speech recognition
German speech to text, based on the Parakeet lineage. About 671 MB. CC BY 4.0.
Optional correction
A German T5 correction model adapted by PJ Labs from oliverguhr/spelling-correction-german-base. About 1.69 GB. Apache 2.0.

The model card documents the original models, adaptation, evaluation and known limitations. The package remains experimental. Its test results do not guarantee transcription or correction quality for your texts.

Download for your Mac

Free early release for Apple Silicon Macs. Models are downloaded separately in the app.

Version
0.1.0 · Build 1
Released
2026-10-03
Download size
36.08 MB
SHA-256 checksum
08761285b861f78e701edf9fdc0cba24768355251738fcb309da22fec9c88903

Verify the downloaded disk image with:

shasum -a 256 ~/Downloads/Vellum-0.1.0-1-macOS-arm64.dmg

Get started

  1. Open the disk image and drag Vellum into Applications.
  2. Start Vellum. Allow microphone access and Accessibility access for insertion into other apps. Follow the permission guidance in Settings.
  3. Download the speech model under Models. Install the correction component only if you want to use it.
  4. Place the cursor in an editable text field. Hold ⌥Z, speak and release the keys to finish.

This early release is locally signed, without an Apple Developer ID or Apple notarization. macOS may block the first launch. After checking the download source, use System Settings → Privacy & Security → Open Anyway if macOS offers it.