Choosing Vito's providers

Vito is free and has no account of its own. Two things happen to what you say — something listens, and something tidies up the words it heard — and for each half you choose who does the work: your own computer, or a provider you hold your own key for. Since 2026.9 neither half needs a key at all.

This page is the map. Every option Vito can use, what it costs, what it asks of you and what it is good and bad at — and, for the four providers that do need a key, a step-by-step guide of its own.

The two halves

The listening half turns your voice into words. Accuracy, the languages you can dictate in and the wait before text appears are decided here. The tidying half turns those words into text worth sending: hesitations gone, sentences mended, line breaks where you asked for them. It is what separates dictation from a transcript of someone talking.

The two are independent. Local listening with cloud tidying is fine, and so is the other way round — pick each half on its own merits.

The listening half

Vito local (Parakeet) Local No usage cost

On your computer · no account, no key · 25 European languages · one-off download of about 900 MB

Vito installs and runs the engine itself: pick it in Settings → Speech recognition, press install once, and that is the whole setup. Underneath it is parakeet.cpp with NVIDIA's Parakeet TDT 0.6B v3; Vito fetches a pinned build and the model, checks both against hashes baked into the release, and runs the engine only while you have that card selected.

• In its favour: nothing to sign up for, nothing to pay, now or ever, and no audio leaves your machine. It is quick, too — on a 12-core desktop a 13-second dictation takes about 1.2 s on CPU and 0.25 s with the Vulkan GPU build. The language is detected for you.
• Against: 25 European languages rather than 60, about 1 GB of memory while it runs and 0.9 GB of disk, and the work lands on your own processor — noticeable on a thin laptop, and on the battery. There is no live transcript: the text arrives when you stop speaking. Keyterms are not used by this engine, though your own corrections still are.

Any OpenAI-compatible endpoint Local or remote

Your own server, or a hosted one · a base URL, and a key only if that endpoint wants one · cost and languages are whatever is behind the URL

Give Vito a base URL and it talks to whatever answers: whisper-server, Speaches or LocalAI on your own machine, or hosted Whisper at Groq and OpenAI. Vito finds whisper-server on its own /inference path, sizes the encoder window to the actual recording — which took a 13-second clip from 9.0 s to 4.1 s on CPU — and joins whisper.cpp's segments back into one line instead of typing its newlines into your document.

• In its favour: the escape hatch. Any engine you like, on hardware you choose, including a machine on your own network doing the work for a laptop that would rather not.
• Against: you are the one running or picking the server, and accuracy, speed and language coverage are its business rather than Vito's. Whisper always encodes 30 seconds of audio, so on a hosted Whisper a two-second dictation can cost what a long one does.

Soniox Recommended

Their servers · your own key · 60 languages · $0.10–0.12 per hour, their list price

The widest language coverage and the cheapest of the cloud engines to run, which is why it is the one we recommend and what the €0.15-per-hour figure on the home page is based on.

• In its favour: all 60 dictation languages, a live transcript, and none of the work on your machine — the reason to pick cloud at all.
• Against: it is prepaid, so a balance has to be there before it does anything, and whether new accounts get trial credit is not something Soniox states on its pricing page. Your audio goes to their servers.

AssemblyAI Free starting credit

Their servers · your own key · $50 in free credits on signup, no card · $0.15–0.45 per hour afterwards

The easiest cloud start: you can have Vito working before you hand over a payment detail, and swapping later is a matter of pasting a different key.

• In its favour: free credit up front, no card asked for, and a second engine to fall back on when a transcript comes out wrong.
• Against: dearer per hour than Soniox once the credit is gone, and your audio goes to their servers.

The tidying half

Whichever you pick, the rules are yours: five built-in sets — Vito default, Verbatim, Business writing, Notes and Messages — and your own on top of them. Vito appends a fixed output contract to whichever set is active, so a rule set can change what cleanup does and can never make the model paste commentary at your cursor.

A local model Local No usage cost

On your computer · an OpenAI-compatible server such as Ollama or LocalAI · no account, no key

• In its favour: free, offline, and nothing you dictate leaves the machine. Paired with local Parakeet it is a Vito with no account anywhere and no bill, ever.
• Against: you need a machine that can hold a decent model, and the cleanup is only as good as the model you gave it. On a modest laptop this is the slowest of the four options.

Groq Recommended Free tier

Their servers · your own key · a genuinely free tier, no card

• In its favour: no cost, no card, and fast enough that the tidying is not what you wait for. By its own terms Groq does not train on what you send, with a zero-retention setting available on top.
• Against: the free tier is capped rather than metered, so a very long day of dictation can meet a limit — and a free tier is the provider's to change. Your text goes to their servers.

Anthropic (Claude) Top-tier

Their servers · your own key · prepaid credits, no free tier

• In its favour: the best of these at turning a transcript into text, and the one to give Vito Assist when you want its spoken answers to be worth reading.
• Against: it costs money from the first dictation, and an account with no balance reports itself as a confusing error rather than an empty account.

Any other OpenAI-compatible provider

A base URL and a key · the provider's own models and prices

• In its favour: if you already pay someone for tokens, Vito can use them. The cost card follows the model you actually selected, so the estimate is not a guess based on the cheapest one.
• Against: you pick the model and live with it. If one spends thousands of tokens deliberating and returns nothing, the fix is the Thinking control, not a bigger token budget.

Three setups that make sense

• Nothing to sign up for. Local Parakeet and a local model. No account anywhere, no usage cost, and neither your audio nor your text leaves the machine. The price is the load on your computer, and 25 languages instead of 60.
• Free, with the work done elsewhere. AssemblyAI's free starting credit for the listening, Groq's free tier for the tidying. Neither asks for a card, your machine stays idle, and you are dictating in ten minutes. The credit does run out eventually.
• The best results. Soniox to listen, Claude to tidy, at roughly €0.15 per hour of speech all in. 60 languages, a live transcript, and the cleanest text of any combination here.

Before you start

A few things worth knowing, whichever way you go:

• You can get all the way to dictating without paying or signing up for anything. That is the local route above. If you would rather the cloud did the work, AssemblyAI's starting credit and Groq's free tier get you there without a card either.
• The paid ones are prepaid. Soniox and Anthropic need a balance before they do anything. A few euros is a lot of dictation, but zero is zero, and an empty account reports itself as a confusing error rather than an empty account.
• If you add a card, set a spending limit while you are in there. It takes a minute and it is the only thing standing between a mistake and your balance. (Free tiers and local models can't run up a bill, so there is nothing to cap while you stay on them.)
• A key is a password. Whoever holds it spends your money — or uses your free allowance. Each guide covers looking after it properly.
• Nothing here is permanent. Every option is a different card in Settings → Speech recognition, and switching is a matter of picking another one. Starting local and adding a key later — or the reverse — costs you nothing.