v1.5.7
Hold it down, and it keeps listening.
Holding the hotkey no longer gets interrupted for no reason. When certain window managers and input methods were running at the same time, the system could clear Option's key state mid-hold, making Voilà think you'd let go — recording would suddenly end, the overlay would disappear, and half a sentence would get cut off. It now treats Tab's physical key state as the source of truth: as long as Tab is still held down, it keeps listening, until both keys are released.
Apple's on-device engine now actually applies custom vocabulary. The new macOS 26 engine now defaults to the same recognition model used by system dictation, so brand names and terms in your vocabulary list are genuinely taken into account; in a Chinese-language context, it also automatically adds contextual phrases for English proper nouns to improve hit rates. Settings let you switch between "Dictation model" and "General model."
Switching engines is cleaner now. When switching between the new engine and the classic engine, the old engine is now fully stopped, so the previous sentence no longer lingers in the overlay. If the classic engine has "On-device only" checked but system dictation isn't enabled, it now explains why in your interface language and offers a one-click jump to the relevant system setting.
Word corrections can now let Apple Intelligence do the work. On macOS 26 with Apple Intelligence enabled, the vocabulary page has a new "Generate with Apple Intelligence" option: give it the correct spelling, and it automatically generates common misheard variants and adds them to the word corrections list; without Apple Intelligence, an offline rule-based version is available instead.
Soniox is more stable. Error handling was aligned with the official docs: when the server explicitly rejects a request, the real reason is now shown directly instead of retrying pointlessly over and over; fixed the vocabulary field for async transcription; and the vocabulary list is now automatically truncated instead of failing outright when it exceeds the server's limit.
The settings window no longer jumps around. The window size is now fixed, so switching tabs doesn't change its height; the connection test and model-list hints in the AI polish config no longer collide on screen.
No settings changes needed — just update and go.