In developmentmacOS 15+ · Apple silicon
Press a shortcut in any app and talk. Speech recognition and the language model both run on your Mac, so VoxtNew works offline, needs no API key and sends nothing anywhere.
30
languages, and 22 Chinese dialects
~0.5 s
until your first words appear
MacBook Air M4, development build
3
features, one shortcut each
0
API keys, accounts or uploads
Features
Transcribe
Press your shortcut in any app, speak, and press it again. The words show while you talk, a language model on your Mac takes out the ums and false starts, and the text is typed into the app in front.
On-device recognition
Qwen3-ASR runs on your Mac's GPU with MLX and writes as you speak. Silero voice detection trims the silence first, so only speech is recognised and a recording without speech is dropped.
Translate
Select text and press the Translate shortcut to translate it in place. With nothing selected, VoxtNew translates what you say. A model made for translation answers in a fraction of a second, between 33 languages.
Rewrite
Select a hasty note, press the Rewrite shortcut and say "make this more polite". With nothing selected, say what to write and the answer goes where your cursor is. Ask for a change, and it carries on as a conversation.
App enhancement
Group apps and websites and give each group a scene: "a team chat: short and casual", "email: complete sentences". VoxtNew matches the app or browser tab in front and adjusts the text to it.
Dictionary
Add the names, products and terms recognition gets wrong, with the ways it tends to mishear them. They go to speech recognition as hints and to the language model as a glossary.
Local models
Instead of a long list, a short one: models we tested and tuned, described by what they are good at and the languages they know. They download once, with resume and checks, and you can see and free the room they take.
Overlay
A small pill near the bottom of the screen shows the feature, a waveform and the words as you say them. It never takes the focus from the app you type into, and the pointer goes straight through it.
Privacy
Speech recognition and the language models run on the Mac. Audio and text are never uploaded, and nothing needs a key or an account.
Open source under Apache-2.0.The source goes public with the first release.
Requirements
Permissions
Microphone
To record while a shortcut is active, and only then. The audio is recognised on your Mac.
Accessibility
To type results into other apps, read the selected text, use fn and single modifier keys as shortcuts and see which app is in front. Without it, results are copied instead.
FAQ
They do the same job in two ways. Auralink sends audio and text to the cloud services you choose, with your own API keys: it stays small and uses the biggest models there are. VoxtNew runs everything on your Mac: it works offline, needs no key and nothing leaves the Mac, but it downloads its models and uses the Mac's memory and GPU.
VoxtNew is in development. Leave your email and we will tell you when the first beta is out.
VoxtNew is in development. Leave your email and we will tell you when the first beta is out.