Voice input

Dictation on Windows – offline, straight into any program

You think faster than you type. Press the hotkey, speak, press it again — the text lands exactly where your cursor is: in Outlook, Word, the browser, your chat window. Speech recognition runs with Whisper or Parakeet on your own machine. Neither the recording nor the text is uploaded.

No account required · Free in your system language

Windows 10 and later · No admin rights needed · No account

Typing is the bottleneck

You know exactly what the email should say. You just need twelve minutes to type it.

Dictation apps exist. Most of them come with three catches: they send your voice to a server, they charge monthly, and they are a separate application — you dictate into it, copy the text, and paste it somewhere else.

Celerando has no window to speak into. The text appears directly in whatever program you are already working in.

Three steps, and no fourth

01
Press the hotkey

Anywhere in Windows, no matter which program is in front. The default is Ctrl+Shift+Space; you can set a different shortcut in Configuration.

02
Speak

Normally. No commands, no pauses in particular places.

03
Press it again

The text is where your cursor was. No copying, no pasting, no switching windows. Prefer hold-and-release? Starting with Business, push-to-talk is available too.

Whisper and Parakeet: 26 languages, all on your PC

Celerando recognises speech with open-source models from OpenAI and NVIDIA. All of them run locally, all are downloaded by the app straight from Hugging Face:

Whisper (OpenAI)

Three model sizes: Base (141 MB), Small (466 MB) and Turbo (574 MB). Runs on the CPU or via Vulkan on any graphics card, optionally via CUDA on NVIDIA cards. Included in every edition, free in your system language.

Parakeet TDT 0.6B v3 (NVIDIA NeMo)

17 European languages with automatic language detection, very low latency, streaming-capable. Runs via Vulkan on any graphics card with 2 GB of graphics memory or more — AMD, Intel or NVIDIA — and otherwise on the CPU. From Business.

What each edition recognises

Free: your system language with Whisper Small or Base. From Pro: all 26 languages, every model, GPU acceleration. From Business: automatic language detection and Parakeet.

  • German
  • English
  • Spanish
  • French
  • Italian
  • Portuguese
  • Dutch
  • Polish
  • Russian
  • Ukrainian
  • Swedish
  • Norwegian
  • Finnish
  • Slovak
  • Catalan
  • Turkish
  • Japanese

Celerando recognises these 17 languages very well with the best model for each, and 9 more well with the occasional correction: Czech, Danish, Greek, Romanian, Croatian, Bulgarian, Hungarian, Korean and Chinese. The basis is the published measurements by OpenAI and NVIDIA. Estonian, Latvian, Lithuanian, Maltese, Slovenian and Arabic can be selected as well, but with clearly more errors. The language is set per dictation or — from Business — detected automatically.

The models: download once, then offline

The models are downloaded once from inside the app and then live in your user profile — no admin rights. Whisper Small is recommended and suggested by the introduction at first start; you pick every other model in the settings.

After that, speech recognition needs no connection at all. Train, plane, basement, isolated company network — dictation keeps working.

ModelDownloadBest for
Whisper Base 141 MBSmallest download, for low-end PCs
Whisper Small (recommended) 466 MBGood accuracy, runs without a graphics card
Whisper Turbo 574 MBBest Whisper accuracy, graphics card recommended
Parakeet TDT 0.6B v3 638 MBFast even without a graphics card, 17 languages detected automatically — from Business

The optional CUDA runtime for NVIDIA cards (about 140 MB) comes from downloads.menterium.net once you select it in the settings. Without it Celerando uses Vulkan — not an error, just a different route to the same graphics card. Which models are offered can change with updates: new, sometimes better ones are added, older ones may be dropped.

Will it run on your machine?

Three tiers, from office laptop to workstation — and none of them strictly needs a graphics card:

Minimum

Windows 10 (build 19041) or Windows 11, 4 GB RAM, a CPU with AVX2 (Intel Core from 2013, AMD Ryzen). Whisper Base and Small run purely on the CPU — even on an older laptop.

Recommended

8 GB RAM and a graphics card with 2 GB of graphics memory or more from AMD, Intel or NVIDIA. Everyday dictation with Whisper Turbo or Parakeet on the graphics card.

Performance

16 GB RAM and a graphics card with 8 GB of graphics memory or more. Long dictations, Smart Modes and the translator on the graphics card.

The Voice input – requirements page has a self-check: three clicks and you know which model fits your machine.

“Menterium” is not a word. To Celerando it is.

Every speech recogniser trips over the same things: personal names, company names, product names, the vocabulary of your profession. You dictate your own surname and get back something that merely sounds similar.

In the Personal Dictionary you enter those words once — together with the variants the recogniser wrongly produces. From then on it sticks.

It works in both directions: the terms are handed to the recogniser before transcription, and wrong spellings are corrected afterwards. With Parakeet (from Business) they also count as hotwords: when a word sounds similar, the recogniser prefers your spelling.

Included from Pro.

From dictation to finished text

After recognition the text passes through several stages, all local, all optional:

Numbers

In German dictation, spoken numbers become digits — amounts, times, decimals and version numbers. In every edition, on by default.

Rules

Spoken symbols become characters (“euro” → €), abbreviations are expanded or shortened, your own replacements via regular expressions. From Pro.

Smart Modes

A local language model (Gemma 3 4B) tidies up: filler words out, punctuation in, soften the tone or phrase it as an email. Usually under a second with a graphics card with 8 GB of graphics memory or more, a few seconds per sentence in the CPU version without one. From Business.

Streaming

The text appears while you are still speaking — piece by piece, directly in the target program. From Business, as is push-to-talk; the two are separate settings.

Your voice never leaves your machine

Cloud dictation means your recording is transferred to somebody else's server, processed there, and the text is sent back.

Anyone dictating client conversations, patient data, personnel files or draft contracts is dealing with professional confidentiality and the GDPR. A cloud service often cannot be approved for that.

Celerando processes recording and text on your PC. What the app contacts at all: the download of each model and additional language you select, the optional CUDA runtime, the update check, license activation including the query of occupied seats, a single installation ping on the very first launch, and the addresses you enter yourself (web app tiles load the site's icon, the connections tile checks your servers). The ping carries a random installation ID, a truncated hash of the device fingerprint, and the app, .NET and Windows versions — no account, no name, no content. License activation and the seat query transmit the computer's hardware fingerprint, which is stored with the activation code. The server does not store the IP address in plain text: a hash of it is deleted after 48 hours, while the approximate location derived from it (country, region, city with an accuracy radius) stays with the record for 24 months. Audio and text are never part of it. Any network monitor will confirm it.

The optional transcription history (from Pro) stores text and recording only in your user profile — off by default.

Celerando and Windows voice typing compared

Windows ships its own voice typing behind Win + H. The difference is not whether it works — it is where.

CelerandoWindows voice typing (Win + H)Cloud dictation apps
Where recognition happens On your PCOn Microsoft's servers (online speech recognition)On the provider's servers
Internet required Never while dictating — only downloads, updates, the first-launch ping and a one-time activation of a paid editionFor every dictationFor every dictation
Recognition model Whisper or Parakeet, model of your choice from ProMicrosoft, not selectableProprietary
Languages 26 well recognised (from Pro), automatic detection from BusinessDepends on installed language packDepends on the service
Your own terms Personal Dictionary, rules, Smart Modes—Depends on the service
Price Free (system language), Pro €36 onceIncluded in WindowsUsually €7–15 per month

As of September 2026. Windows voice typing is perfectly fine for occasional notes — the only question is whether your dictation is allowed on someone else's server.

When typing is not an option

When your hands stop cooperating

Tenosynovitis, RSI, tendinitis. Dictation is then not a convenience but the condition for continuing to work at all.

When you write in a foreign language

Speaking is almost always easier than writing.

When you produce text all day

Emails, minutes, documentation, notes.

Pay once. Not every month.

A dictation subscription at €5 a month sounds like very little. After one year that is €60, after three years €180.

Celerando Pro costs €36 once. Updates within the version are included.

Frequently asked questions about offline dictation

Is Celerando dictation software?

Yes: dictation software for Windows that works fully offline. You press a hotkey and speak, and the text appears in the program you are writing in; there is no separate dictation window. Speech recognition runs on your PC with Whisper or, from Business, with Parakeet. Dictation is included in the free version (30 seconds per recording, your system language). Celerando also comes with a local translator, text recognition (OCR, from Pro) and QuickShare.

Does dictation work without internet?

Yes, fully. The speech model is downloaded once from inside the app — Whisper Small at 466 MB is recommended, Base is 141 MB, Turbo 574 MB and Parakeet is 638 MB — and then lives in your user profile. Recognition itself opens no connection; try it in flight mode.

Do I need an account, and does Celerando check the license online regularly?

No to both. The free version needs no license. You activate a paid edition once with the code from your purchase email; the PC's hardware fingerprint is stored with the code. After that, a signed license file sits on your PC and Celerando checks it locally, with no expiry date. Only when you open the license page in the settings does the app ask for the occupied seats and show them for multi-seat licenses; without internet that line stays empty and the license stays valid. PCs that must never go online are activated with a license file (see “Activation without an internet connection”).

Which languages are recognised?

26 languages are recognised well. Very well: German, English, Spanish, French, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Swedish, Norwegian, Finnish, Slovak, Catalan, Turkish and Japanese. Well, with the occasional correction: Czech, Danish, Greek, Romanian, Croatian, Bulgarian, Hungarian, Korean and Chinese. The basis is the published measurements by OpenAI and NVIDIA with the best model for each language. Your system language is free, all of them are available from Pro, and from Business Celerando detects the language automatically. Estonian, Latvian, Lithuanian, Maltese, Slovenian and Arabic can be selected as well, but with clearly more errors.

Whisper or Parakeet — which should I use?

Whisper is the default: part of every edition, in all 26 languages from Pro, and very accurate with the Turbo model on a graphics card. Parakeet pays off if you dictate in a European language, use a lot of technical terms or don't have a strong graphics card: the words from your Personal Dictionary feed straight into the recognition as hotwords, and Parakeet is fast on the CPU too. Parakeet is included from Business.

Is the recognition as good as with cloud services?

No. Whisper and Parakeet run on your PC and are considerably smaller than the large cloud models; they do not reach their quality. You notice it most with rare technical terms, dialect and background noise – for technical terms the Personal Dictionary helps. For clearly spoken emails, notes and texts the recognition works well; try it yourself with the free version.

How big is the download, and where are the models stored?

Whisper Base is 141 MB, Small 466 MB and Turbo 574 MB, Parakeet takes 638 MB, and the optional CUDA runtime adds about 140 MB. The models come straight from Hugging Face and live in your Windows user profile under %LOCALAPPDATA%\Menterium\Celerando — no admin rights, and they can be deleted from inside the app.

Do I need a graphics card?

No. Whisper Base and Small run on the CPU, even on an older office PC, and so does Parakeet. A graphics card with 2 GB of graphics memory or more makes Whisper Turbo and Parakeet much faster — driven via Vulkan, optionally via CUDA on NVIDIA cards. GPU acceleration is included from Pro. If you select CUDA without the runtime installed, Celerando falls back to Vulkan — no error, no abort. Speech recognition needs a CPU with AVX2 (Intel Core from 2013, AMD Ryzen).

Is my voice stored anywhere?

Not outside your machine. The optional transcription history (from Pro) keeps text and recording in your user profile so a dictation is not lost when it lands in the wrong window — it is off by default and can be cleared at any time.

In which programs does dictation work?

In all that accept text input — Outlook, Word, browsers, Teams, Slack, editors, forms. Celerando inserts the text via the clipboard and Ctrl+V by default; for remote desktop and terminals there is a mode that types character by character.

How long can a recording be?

In the free version 30 seconds per recording. From Pro, unlimited.

Does the text appear while I am still speaking?

That is streaming mode, included from Business — the text appears piece by piece as you speak. In Free and Pro the text is inserted as soon as you stop the recording.

What are Smart Modes?

A local language model that post-processes the recognised text — removing filler words, adding punctuation, softening the tone, phrasing it as an email. Celerando uses Google's Gemma 3 4B for this. With a graphics card with 8 GB of graphics memory or more (AMD, Intel or NVIDIA) it runs on the card alongside speech recognition, usually in under a second; without one, you pick the more compact CPU version of the same model, which takes a few seconds per sentence. From Business.

Is it suitable for law firms, medical practices and businesses bound to confidentiality?

The starting point differs from a cloud service: there is no third party processing recording or text, and so no data processor that would have to be bound by contract. Whether that is sufficient for your case is for your data-protection review to decide — this page provides the technical facts for it.

Do I need admin rights?

No. Celerando installs into your user profile, and so do the models.