OpenQuack 2.0.0-alpha.17

Highlights

Non-English & mixed-language dictation now works on auto-detect.

Before this build, with language set to auto-detect, OpenQuack silently translated non-English speech into English instead of transcribing it (a misconfiguration that disabled WhisperKit's language detection entirely). Now it detects the spoken language and transcribes it correctly.

Measured on medium (M4, 16 GB):

Case	Before	After
Multilingual (zh/ja/ko/es/fr/de, auto)	253% WER / 156% CER	16.7% / 3.7%
Mandarin clip → output	English translation	correct Chinese
Mixed EN/ZH (code-switch)	translated to English	keeps Chinese + inline English terms (e.g. 我需要 refactor 这个 function)
English (LibriSpeech, short)	2.6%	2.6% — unchanged

The fix is zero-cost when you pin a language in Settings, and adds only a single in-decoder detection pass on the auto path (no extra audio encode). Model size unchanged.

Fixes #17. See SPEC-021 and PR #63.

Install

brew tap larryxiao/openquack https://github.com/larryxiao/openquack
brew install --cask openquack

Or download the DMG below. This alpha is unsigned (like its predecessors); on first launch, right-click → Open, or xattr -dr com.apple.quarantine /Applications/OpenQuack.app.

Known follow-ups

• CJK↔Latin spacing in mixed output (refactor这个 → refactor 这个)
• Simplified/Traditional consistency on longer Chinese