Skip to content

VoxKey 0.3.0

Choose a tag to compare

View all tags
@rodrigouroz rodrigouroz released this 10 Sep 18:28
· 11 commits to main since this release
v0.3.0
dbaa9c9

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.

Assets 5
Loading