Privoscribe, explained.

1. 1

   Download

   The setup file from this website – for Privoscribe or Privoscribe AI. It is small: it fetches the actual installer itself.

2. 2

   Start

   On the first start Windows may warn: “Windows protected your PC”. Click More info, then Run anyway. The setup file downloads the installer, checks it and starts it.

3. 3

   Set up

   Folder, shortcuts, “Start with Windows” – for Privoscribe AI also the AI model. Then the installer fetches the parts and the app starts.

Good to know

The setup file

It fetches the current installer (70 MB) from privoscribe.com and starts only one that Privoscribe has approved with its own key. A changed or foreign file is deleted without being started. Without internet it points you to the offline package.

What the installer fetches

Speech recognition with its model (0.7 GB), with an NVIDIA graphics card its runtime (0.8 GB), for Privoscribe AI also llama.cpp and the AI model. All straight from the official sources – GitHub, PyPI and Hugging Face – every file verified by checksum. An interrupted download continues where it stopped.

Offline package

A ZIP with everything: unpack it, start “Privoscribe Setup.exe” – nothing is downloaded. Privoscribe AI comes with the fast model (E2B); fetch the more accurate one in the app if you want it.

Where to

For your Windows user, without administrator rights, into %LOCALAPPDATA%\Programs\Privoscribe or …\Privoscribe AI.

Updates

About once a day the app asks whether there is a new version (you can switch this off). If there is, it shows in the icon’s menu. Then simply run the setup file again or download it anew – it always fetches the latest version. Your settings stay, and only what changed is downloaded.

Uninstall

Windows Settings → Apps → Privoscribe → Uninstall.

Both editions?

They listen to the same shortcut, so only one runs at a time. Privoscribe AI offers to close Privoscribe when it starts.

1. 1

   Hold

   Hold Ctrl+Win. The pill appears at the bottom centre: Privoscribe is listening.

2. 2

   Speak

   Just talk, pauses and “um” included. Speech recognition adds punctuation and capitals itself.

3. 3

   Let go

   A moment later the text is at your cursor.

The first start

Once started, Privoscribe sits quietly in the notification area at the bottom right and waits for your shortcut. From then on it starts with Windows – you can switch that off in the settings.

Good to know

Hands-free

While holding, press Space, then let go – Privoscribe keeps listening. Finish with Ctrl+Win or the red button.

Discard

Esc or the ✕ in the pill. Nothing is inserted.

Your own shortcut

Any combination works, such as Ctrl+Space, F8 or mouse button 4 – with “Hold” or “Toggle” (press once to start, again to stop).

Dictionary

If a name always comes out wrong, add the fix: “Frau Meyer → Frau Maier”. Whole words, case doesn’t matter.

Microphone

The Windows default or a particular device. If that one isn’t plugged in, Privoscribe uses the default.

No text?

Some programs run with administrator rights and don’t accept input from normal programs. Use “Copy last text” in the icon’s menu – and Ctrl+V.

Two shortcuts

Ctrl+Win	plain dictation – exactly like Privoscribe, no AI
Ctrl+Alt+Win	write with AI – the pill turns blue

Already holding Ctrl+Win? Add Alt and the same recording becomes an AI recording. Prefer a single shortcut? Choose “One shortcut” in the settings; then “Write with AI” in the icon’s menu decides whether Ctrl+Win dictates or writes.

First check …

When you let go, a card appears next to your cursor with what Privoscribe heard. Fix any mishearing right in the text.

1. 1Preset and output language – what it should become, and in which language.
2. 2Instructions – on: “Write my landlord …” is a request. Off: everything you said is the text.
3. 3Your transcript – editable. Shift+Enter starts a new line.
4. 4Continue with Enter – or insert it Without AI.

1234

… then insert

Gemma rewrites and translates in the same pass. On a graphics card that usually takes about a second.

1. 1Change preset or language – Ctrl+R then writes it again.
2. 2Fact check – here the AI turned 30 people into 32. The card notices.
3. 3Refine in one click: shorter, more detail, friendlier, more formal.
4. 4Or say what should change – type it, or dictate into it with Ctrl+Win.
5. 5Insert with Ctrl+Enter. “Original” inserts your transcript instead.

12345

Refine it – three ways

Say it

Hold the AI shortcut while the card is open and say, for example, “make it shorter” or “add that I’m free on Tuesday”.

Buttons

Shorter · More detail · Friendlier · More formal.

Field

Type and press Enter – or click into it and dictate with Ctrl+Win.

Each change takes about a second; the back arrow brings the previous version back. Whatever you added yourself doesn’t count as invented in the fact check. With “More detail” the small model likes to embellish a little – give it a quick read.

The fact check

If the AI changes or drops a number, a date, a time, an amount, a percentage, an email address, a link or a phone number – or deletes an “I think” – the card marks it in yellow. Across languages, too: “14:30” is “2:30 PM”, “Donnerstag” is “Thursday”. Your transcript is always kept.

Write it for me

Say what you want instead of dictating the text: “Write my landlord that the heating has been broken for a week and he should send someone.” You get a finished letter with greeting, request, thanks and sign-off. If a dictation starts with “Write …”, “Reply …” or “Draft …”, the card picks this preset on its own and tells you.

The facts come only from you. Anything missing is written around – or, if you switch on “Placeholders for missing details”, becomes a [placeholder] to fill in.

Reply

Select and copy a message (Ctrl+C), click into the reply box, hold the AI shortcut and say what you want to answer: “Say yes, but only from 3 pm.” The reply comes in the message’s language and matches its tone.

For this, Privoscribe AI keeps only the text you copied last – for three minutes, in memory. It ignores passwords copied from password managers.

Emails: greeting and sign-off

Privoscribe AI sets the tone with fixed rules, from what you said – not the model:

If you end with “Cheers, Lena”, the name goes under the sign-off – even after “Shorter” and friends. German emails get the matching German forms, from Sehr geehrter Herr … to Viele Grüße, with the right formal or informal “you”.

To see the same words as a note, a message or an email, turn the preset dial on the home page – with real outputs.

Your own presets

Under Settings → Presets you give a preset of your own a name and an instruction, such as “Turn my dictation into concise documentation for software development.” The basic rules always apply: invent nothing, keep numbers, names and dates unchanged.

Output language

“Same as input” (the default), German, English, French, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Swedish, Danish, Norwegian, Finnish, Romanian, Turkish, Ukrainian, Russian, Arabic, Chinese, Japanese, Korean, Hindi – or “Other language …”. Gemma isn’t equally good at all of them.

Speak German, preset Email, output English: you get an English email. “Translate” with “Same as input” turns German into English and English into German.

Languages you can speak

Speech recognition understands 25 European languages and notices by itself which one you speak: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian.

The models – and why these

You download the model once in Privoscribe AI, under Settings → AI model – verified by checksum, resumable and removable. Plain dictation works without it.

Why two models?

Listening and writing are different jobs. A small model that only listens is faster and more accurate than a big one that is meant to do everything – and Privoscribe without AI needs only this one.

Graphics card or processor

With an NVIDIA graphics card both models run on CUDA. On AMD and Intel, Privoscribe tries Vulkan, otherwise the processor – that always works, just slower. “Automatic” switches to the processor when a game or a 3D program is using the graphics memory.

How fast is it?

That depends on your PC. On a graphics card a sentence is usually recognised in a fraction of a second and rewritten in about a second. On the processor, rewriting takes a few seconds, and noticeably longer with E4B. The times on the home page were measured on an NVIDIA GeForce RTX 2080 Ti.

Quiet in the background

While Privoscribe waits for your shortcut, it does practically nothing. Gemma is only loaded once you start speaking with AI – alongside the recording – and frees its memory after 15 minutes without use (adjustable). Gemma runs as a separate process: if it ever crashes, Privoscribe AI carries on, and your transcript is still there.

Recordings

in memory only – never saved as a file, never sent anywhere.

Texts

go into your program, nowhere else.

Account, cloud, telemetry

there are none.

Logs

for troubleshooting: times and lengths, never your words.

Clipboard

For replies, Privoscribe AI keeps the text you copied last for three minutes, in memory only.

Internet

to install (the offline package not even for that), in Privoscribe AI for the AI model – straight from Hugging Face, when you start the download – and about once a day to ask for a new version, as long as “Check for updates” is on. Nothing about you or your texts is sent.

Where things are

System

Windows 10 or 11, 64-bit

Graphics card

not required. Much faster with NVIDIA; Gemma 4 E4B needs 6 GB of graphics memory.

Memory

8 GB. For Privoscribe AI without a graphics card, 16 GB is better.

Disk space

Privoscribe up to 1.8 GB · Privoscribe AI up to 2.2 GB plus 3.3 GB (E2B) or 5.2 GB (E4B)

Microphone

any that Windows knows

Internet

only to install – not at all with the offline package

Your question isn’t here, something doesn’t work or you have an idea? Write to me →