VoxKey vs
Apple Dictation.

What actually differs?

Offline use alone does not distinguish them: Apple Dictation can also process speech on-device. Check its Keyboard settings for the processing mode available on your Mac.

Apple sources: Dictation guide, spoken commands, and language availability. VoxKey sources: 0.6.0 usage guide and implementation scope.

A native Mac app, from menu to models

VoxKey is built with Swift and AppKit. Its menu bar controls and Settings window use native macOS components, and its speech and editing models run inside the app. There is no local server, Python environment, or Homebrew runtime for you to set up.

Apple Dictation is native to macOS too. VoxKey’s reason to try it is the combination of selectable speech models, deliberate hold-to-talk delivery, and optional local editing. You can inspect the implementation in the open-source app.

English cleanup that stays on your Mac

Improve transcription is an optional step that removes fillers and repetitions and improves English grammar. It defaults off. When enabled, three stages turn speech into the text delivered to your editor:

1. Whisper transcribes. The selected local speech model converts your recording to text.
2. S1-mini cleans up. Superwhisper’s editing model runs through directly linked llama.cpp with Metal support. Prompt lookup accelerates decoding by reusing matching text from the input.
3. GECToR checks grammar. A Core ML grammar model applies restricted edits. Guards protect technical literals and limit word substitutions in this stage.

VoxKey integrates and validates these models; it did not create S1-mini or GECToR. Failed, cancelled, unsupported, or oversized improvement requests preserve the original recognition result.

Improvement requires English as the selected language and up to 2.03 GB of optional model downloads. It pauses for Spanish, other languages, and Automatic. The GECToR model is for noncommercial use, under terms separate from VoxKey’s MIT-licensed app. Models can still change names, meaning, or spoken corrections; review important text.

Check offline use before you need it.

Apple’s Dictation guide explains how to check whether general text dictation is processed on-device; its notice distinguishes that from dictation in a search box. Read the privacy information for your configuration.

VoxKey downloads its speech model during setup, then uses that model locally. It does not save recordings or a dictation history. Once text is delivered or copied, the destination app and clipboard determine what happens to it.

Before relying on either option while travelling, complete setup and try a harmless sentence in a local document without an internet connection. Successful downloading alone does not prove that your chosen setup works offline.

Dictating in English and Spanish

In VoxKey, open Settings → Models & Languages, download a Whisper v3 Turbo model, and select Use Model. Choose Spanish or English explicitly for a short phrase, or try Automatic for language detection. An English-only Distil model will not provide Spanish dictation.

Spanish dictation works independently of the optional English improvement pipeline. Selecting Spanish or Automatic pauses improvement, while speech recognition remains available.

A small published VoxKey model experiment used synthetic English and Spanish speech. It did not test Apple Dictation, personal voices, or end-to-end text delivery, so it cannot establish that VoxKey is more accurate or faster.

A small test with your own voice

Use the same microphone, room, language, and text for both apps. Note your Mac chip, macOS version, VoxKey version and model, and whether transcription improvement is enabled. Start with improvement off to compare ordinary dictation.

These are suggested prompts, not recorded results.

1. Check the words. Repeat each passage a few times. Note missing words, names, accents, and any change in meaning. Score punctuation separately.
2. Check the wait. Time from your last spoken word to usable text. For VoxKey, include releasing the trigger and waiting for insertion. Decoder-only timings are not the same measurement.
3. Check your apps. Try a local document and the editors you use each day, with disposable text. Check insertion at the cursor and replacement of selected text. Do not send or submit the result.
4. Check the controls. Try starting, finishing, and cancelling. If spoken punctuation or paragraph commands matter to you, test those separately.

VoxKey’s editor compatibility is still being tested. If delivery cannot be confirmed, Safety Net makes the last result available for recovery; it is not a saved history. The 0.6.0 release also notes an unresolved intermittent Right Command activation issue. See the release limitations.