VoxKey 0.4.0

VoxKey 0.4.0

First-time setup now makes your spoken language, dictation key, and model download
status easier to find. Requires Apple Silicon and macOS 26 or later. Internal
build number: 11.

Set up dictation your way

• Choose your spoken language before activating your first model. The choice stays
  selected while the model downloads.
• Language and trigger-key controls appear in onboarding before the first dictation
  test. Choose Globe/Fn, Right Option, Right Command, or Right Control.
• English-only models explain that Spanish and other languages require Whisper v3
  Turbo. The model chooser prevents activating an English-only model with a
  non-English language selected.

Clearer download feedback

• An animated starting indicator appears immediately, including while the downloader
  looks up the model files.
• Setup and Models and Languages show download progress and elapsed time. The menu
  also reports an active model download.
• Completed transfers show a finishing state until the downloaded model is checked.
• Overall time remaining is unavailable: the downloader measures progress by file,
  rather than by byte, so the percentage cannot support a reliable countdown.

Menu alignment

Menu labels now align consistently across sections. The automatically inserted
Settings gear no longer shifts the upper section to the right.

Validation and limits

Automated tests cover language selection before activation, onboarding language
and shortcut controls, download startup, progress, completion and retry, and menu
alignment. Light and dark layouts were rendered and inspected.

Live first-install dictation on another computer and an end-to-end Sparkle update
have not been re-tested for this release. Editor compatibility still depends on
the destination application's macOS Accessibility support.

VoxKey 0.3.1

VoxKey 0.3.1

Public builds now exclude VoxKey diagnostic logging and inspection commands. The
website is available in English and Spanish. Requires Apple Silicon and macOS 26
or later. Internal build number: 10.

Diagnostics stay in internal builds

• VoxKey diagnostic messages, logger setup, and inspection entrypoints are compiled
  only into explicitly internal builds. Public app, preview, and distribution
  builds exclude them, regardless of optimization level.
• Packaging checks the emitted executable for diagnostic markers, so an accidental
  internal build flag cannot silently produce a public diagnostic build.
• Sparkle system profiling is explicitly disabled at startup. VoxKey has no
  telemetry transport or analytics/crash-reporting SDK.
• These changes apply to VoxKey's own diagnostics. macOS and dependencies can still
  emit system logs; this release does not promise zero logging across the process.

An English and Spanish website

• The public website has English and Spanish pages with a language switcher.
• Both pages describe model selection, multilingual dictation, privacy, installation,
  and the English-only grammar correction option, and link to the same release.
• This translates the website; it does not translate the native app interface.

Validation and limits

All 233 app tests and nine artifact-privacy regression tests passed. Packaged
grammar controls were exercised, and all 28 VoxKey diagnostic markers were absent
from the public app. The downloaded
release is checked against its checksum, signing certificate, notarization ticket,
and Gatekeeper assessment before publication.

Editor compatibility still depends on the destination application's macOS
Accessibility support.

VoxKey 0.3.0

VoxKey 0.3.0

VoxKey now has a model library with four cards, multilingual dictation, and a
refreshed native interface. Requires Apple Silicon and macOS 26 or later.
Internal build number: 9.

Choose the model that works for your voice

• Compare compact and full versions of Distil-Whisper v3 and Whisper v3 Turbo.
  Every card explains language support, download size, and the model's tradeoffs.
• Download several models and keep them on your Mac. Downloading is separate from
  activation: choose Use Model when you want to switch. One model is active;
  existing downloads remain available for future use.
• Open the library with Choose Model… during setup, Manage Models… in
  Settings, or Models and Languages… in the menu bar.
• Model and language choices persist. Finish dictation before switching;
  preparation failure preserves the previous working model and language.

Dictate in Spanish and other languages

• Both Turbo variants support Spanish and other Whisper languages, with explicit
  language selection or automatic detection. Distil-v3 supports English only.
• Batch and streaming transcription honor the selected language. Text delivery
  preserves Spanish opening punctuation, and word replacement uses local language
  detection instead of forcing English.
• Optional grammar correction remains English-only. It is paused for Spanish,
  other non-English languages, and Automatic. The saved preference returns when
  you select English again. The grammar model retains its noncommercial-use terms.

A consistent native design

• A lime keycap on forest green replaces the old app icon. The menu bar uses a
  monochrome version while ready.
• First-run setup and everyday Settings have separate windows. Settings contains
  model access, trigger, capture mode, microphone, grammar, login, and feedback.
• Shared cards, native controls, system colors, serif titles, and consistent
  spacing connect the model library, onboarding, Settings, Vocabulary, and Safety Net.
• The public page shares the brand mark and typography, adds dark mode, and keeps
  installation steps visible.

Validation and known limits

All 233 local app tests and nine artifact-privacy regression tests passed.
The local test suite covers the model cards, file inventory, saved selection,
language restrictions, and shared grammar controls. Real Core ML inference checks
Spanish through batch and streaming paths and verifies failed model replacement.
Rendered UI checks include light and dark appearances. Distribution retains the
artifact privacy checks introduced in 0.2.1.

All five research candidates, including a local Distil-v3.5 conversion, were
benchmarked on synthetic speech.
That small sample does not predict accuracy for an individual voice. There is no universal recommended
model or accuracy guarantee. Distil-v3.5 is not a downloadable library option.

Spanish grammar candidates did not meet the required quality and supporting-evidence
bar, so Spanish correction is excluded. Microphone, accent, long-form dictation,
and editor compatibility still need testing in the apps you use.

VoxKey 0.2.1

VoxKey 0.2.1

This release updates the speech engine, unifies the app's appearance, and restores
packaged downloads with stronger build-privacy checks.

• Upgrade to WhisperKit 1.1.0 while keeping the existing local Distil-Whisper models.
• Use the upstream vocabulary prompt fix and a smaller dependency graph.
• Share the forest-green palette and typography across Settings, Vocabulary,
  Safety Net, and the recording overlay, with matching website styling.
• Remove developer build paths from packaged executables and verify the final
  disk image, bundled resources, and release metadata before distribution.

Requires Apple Silicon and macOS 26 or later. Internal build: 8.
Optional grammar correction remains off by default and uses a separately
downloaded model licensed for noncommercial use.

Validation and limits

The automated Swift suite and packaging-privacy regression tests pass. Packaging
checks cover bundled grammar resources and controls, signatures, and final-image
privacy. The app is Developer ID signed, notarized by Apple, and stapled; the DMG
is an unsigned container containing that verified app.

The speech-engine upgrade received maintainer dictation testing before release.
Renewed physical microphone and editor testing of the combined release is still
needed. Report problems with made-up text and never attach private dictation or
vocabulary.