Powerful Features.
Zero Bloat.
Offline speech recognition on your own CPU, text typed straight into the app you work in, and optional AI when you want it.
Live Typing
Dictate and watch the text type itself into your app as you speak; the last words land when you stop. Prefer the classic way? The regular end-of-recording mode is always there. Live typing is part of Pro.
Global Hotkey Control
Press your configured global hotkey anywhere in Windows to start/stop dictation, hold it down like a walkie-talkie (push-to-talk), or let auto mode tell a tap from a hold. No need to switch windows — the app floats on top of your work. Push-to-talk and auto mode are part of Pro.
Private by Default
Your speech is transcribed locally by Parakeet Ultra, an improved version of NVIDIA's Parakeet model, running on your CPU. Nothing is sent to the cloud unless you explicitly turn on AI features. Dictation keeps working with the internet switched off.
Multilingual, Offline
Parakeet transcribes 25 European languages entirely on your device and works out by itself which one you are speaking — no language to pick first. Norwegian (Bokmål & Nynorsk) comes through an optional NB-Whisper pack you download when you need it, and with AI turned on your dictation can be translated into almost any language. The app itself is available in 14 languages.
AI: Translate & Rewrite
Translate your dictation into almost any language, select text anywhere and rewrite it by voice — or select nothing, and your dictation arrives written clean. AI is part of Pro: use the built-in AI with nothing to set up (Pro + AI), or bring your own key for OpenAI, Anthropic, Google Gemini, DeepSeek, Groq, Mistral, OpenRouter, YandexGPT or any OpenAI-compatible endpoint; your key stays on your machine. Only the text is sent, never the audio.
Built for Workflow
Custom hotkeys (keyboard chords, a double-tap modifier, or mouse buttons), a full audio panel with mic picker, sensitivity and a live level meter, noise reduction, a searchable transcription history, and an on-demand Norwegian language pack. The transcription history is part of Pro.
Smart Injection
VibeVox types text directly into your active application. If you switch windows mid-dictation, the text lands safely on your clipboard instead of the wrong app. Pasted dictations stay out of Windows clipboard history (Win+V). Fine-tune the output with formatting toggles (Pro) — like a lowercase first letter or trimming the final period.
Hands-free & long takes
Pro adds hands-free dictation: switch listening on with your hotkey, and each phrase starts when the camera sees your lips move and the microphone hears you — the words are still recognised from the audio, not from the video. Plus takes of up to 5 minutes, Auto Enter that sends a message the moment it is typed, and music and videos muted while you dictate.
Noise suppression
Two levels for a noisy room: the normal one removes a steady hum such as a fan or air conditioning, and AI noise suppression handles a café or the street. Both run on your computer. Off by default, because a quiet room needs neither; live typing is off while it runs.
Your own words and terms
Names, terms and abbreviations that recognition gets wrong: write the right spelling once, and close-sounding mistakes turn into it. You can also replace a spoken word with anything you like, such as your phone number or address. Part of Pro.
Works with NVDA and Narrator
NVDA and Windows Narrator read the Settings window aloud — every switch says its name, whether it is on, and what it does. A short sound marks the start and the end of each recording, so you know it is listening without looking at the screen.
AI is optional
Dictation runs on your computer and needs neither AI nor the internet. AI is an add-on for three jobs: translating what you dictate, rewriting selected text by voice, and writing your dictation out clean. You switch it on yourself, and only text is sent, never your voice.
Your own key
Already have an API key from OpenAI, Anthropic (Claude), Google Gemini, DeepSeek, Groq, Mistral, OpenRouter or YandexGPT? Paste it into Settings. The text goes straight from your computer to your provider, and you pay that provider at its own rates. Any OpenAI-compatible service works too. The key stays on your computer.
Included in Pro (€29, one payment)
Built-in AI
Nothing to set up: no keys, no accounts with anyone else. Switch it on and it works, and we pick the model for you. The text is processed in Europe.
Pro + AI subscription, €5.99 a month, cancel anytime
Don't need AI? Leave it off: everything else works without it. During the first 14 days you can try both ways.
25 languages, fully offline
Parakeet auto-detects the language and transcribes it entirely on your device. No cloud, no account — your voice never leaves the machine.
Optional packNorwegian? Also covered.
Bokmål and Nynorsk run through NB-Whisper Small — a model by the National Library of Norway. VibeVox downloads it once (~670 MB) right from Settings, and it works fully offline too. To pick it: Settings → General → Recognition Model → Norwegian.
Technical Specifications
Parakeet Ultra is an improved version of NVIDIA's Parakeet TDT 0.6B v3 speech model: retrained for accuracy and packed by us into 8-bit so it runs fast on an ordinary processor. On the FLEURS benchmark across 25 languages it makes about half as many word errors as the original model (11% vs 20%). Norwegian has its own model: NB-Whisper Small from the National Library of Norway, trained on Norwegian speech.
Under the Hood: Next-Gen Architecture
Token-and-Duration Transducer (TDT)
Unlike traditional RNN-Ts that predict one token at a time, Parakeet's TDT architecture jointly predicts both the token and its duration. This allows the model to skip silence and redundant frames during inference, drastically reducing compute time.
- Joint Token & Duration Prediction
- Skip-Frame Inference Acceleration
FastConformer with 8x Subsampling
Built on the Conformer architecture but optimized for speed. It uses an 8x downsampling factor to process audio frames more efficiently, combining the global context of Transformers with the local feature extraction of CNNs.
- 8x Downsampling Factor
- Runs entirely on CPU — no GPU required
Start dictating
for free
Every paid feature is unlocked for your first 14 days. No card, no account.