Говоріть.
Vito надрукує це — чисто.

🎙️ Dictation without an account, on your own machine

Vito can now install and run a speech engine itself. Pick Vito local (Parakeet) in Settings → Speech recognition, press install once, and that is the whole setup: no provider account, no API key, no server of your own, and no audio leaving your computer.

Behind it is parakeet.cpp with NVIDIA's Parakeet TDT 0.6B v3 — 25 European languages, detected automatically. Vito downloads a pinned build and the model (≈ 900 MB, once), checks both against hashes baked into the release, and runs the engine as a background process only while you have that card selected.

Measured on a 12-core desktop:

It needs about 1 GB of RAM while it runs and 0.9 GB of disk. There is no live transcript — the text arrives when you stop speaking. Keyterms are not used by this engine; your corrections still are.

🔌 Or point Vito at any OpenAI-compatible endpoint

The other new provider takes a base URL and talks to whatever is behind it: whisper-server, Speaches, LocalAI on your own machine, or hosted Whisper at Groq and OpenAI.

Whisper always encodes 30 seconds of audio, so a short dictation costs the same as a long one. Where whisper.cpp may be listening, Vito now sizes the encoder window to the recording — that took a 13-second clip from 9.0 s to 4.1 s on CPU. It also finds whisper-server on its own /inference path, and joins whisper.cpp's segments back into one line instead of typing its newlines into your document.

✍️ Cleanup rules you can write yourself

The cleanup prompt is the biggest lever on what comes out, so it is yours now. Five built-in sets — Vito default, Verbatim, Business writing, Notes, Messages — are always in the list and read-only; start your own from any of them.

Only the *rules* are yours: Vito appends a fixed output contract to whichever set is active, so a rule set can change what cleanup does and can never make it paste commentary at your cursor.

New Thinking control alongside it, for endpoints that support reasoning_effort. Turning it off is the fix when a model spends thousands of tokens deliberating and returns nothing — raising the token budget never was.

🧾 Why a dictation came out unpolished

When the AI pass fails, the raw transcript is pasted and a toast disappears seconds later. The reason is now kept on the history entry, so you can still find out hours afterwards. The notification points you there instead of truncating a provider error into uselessness.

The out-of-credit card can be dismissed, and stays gone until that account actually runs dry again.

🎛️ The speech settings, rearranged

The page was one long list of every field of every provider. Each card now carries its own fold-out — open for the provider you use, closed for the rest — with the provider's logo and one line saying what it asks of you: an account somewhere, a server of your own, or nothing but a download.

🐛 Fixes

• A dropped streaming connection fell back to AssemblyAI's async API whatever provider you had chosen. A Soniox user without an AssemblyAI key simply lost the dictation.
• The silence watchdog and auto-stop read speech activity off the live transcript, which a batch provider never sends. They now fall back to the microphone level, so both work with the new providers.

🌍 Translations

All of the above is translated in all 60 languages.

Install

• Windows — Vito-Setup-2026.9.exe (per-user, no admin).
• macOS — Vito-2026.9.dmg — drag Vito to Applications. First launch needs right-click → Open, then two permissions (Microphone, Accessibility).
• Linux — Vito-2026.9-x86_64.AppImage — chmod +x and run — or the Flatpak from vinceecniv.github.io/Vito.

Checksums (SHA-256)

b9ed8de7e58fb36c18e1896e45098055f4509a3b38525108dd7d6ac6e02d3f5c  Vito-Setup-2026.9.exe
cdca76fc98f3d7880165ad95a215d82a6f0011f570e2ed58c40d9a5785a08b14  Vito-2026.9.dmg
e188d28c1d570fabaa31f4498bcb8ce74238ae64e07e83ec534f5ee9d33aeac4  Vito-2026.9-x86_64.AppImage

Free, open-source, bring-your-own-key — no subscription. Setup guides at vito.talk.