Dictation
Dictate anywhere on your system with a global hotkey, have the finished text inserted where you were typing, and configure every part of it
Dictation turns speech into finished text in whatever application you happen to be in. You press the hotkey, speak, and the text lands in the field you were typing in: an email, a ticket, a chat message. The DeutschlandGPT window never has to come to the foreground.
Dictation exists only in the desktop app, because it needs native microphone capture and a system-wide hotkey. It also has to be enabled for your account. If you do not see a Dictation entry in the settings rail, ask your administrator.
Where to find it
| What | Where |
|---|---|
| The hotkey and how it fires | Settings, Shortcuts, group "System-wide" |
| Everything else | Settings, Dictation |
| Your past dictations | Settings, Dictation, at the bottom under "History" |
Open settings with Cmd/Ctrl+,. The hotkey deliberately lives with all the other shortcuts rather than in the dictation panel, because it describes a key press and not what happens to the text.
Dictating
Put the cursor where the text should go, then press the dictation hotkey. The default is
Ctrl+Alt+D on both macOS and Windows (on macOS that is the Control key, not Cmd). A small
overlay appears and shows that it is listening.
Speak. How you stop depends on the trigger setting: press the hotkey again in toggle mode,
or release the keys in push to talk mode. Esc cancels.
The recording is transcribed, refined according to the selected mode, and inserted at the cursor.
If the text cannot be inserted automatically, most often because the Accessibility permission is missing on macOS, it is placed on the clipboard instead and the overlay says so. Dictation never ends with no text at all.
Permissions
The two rows at the top of the dictation settings are not preferences, they are preconditions. Both are granted in your system settings, and the panel links straight to the right pane.
| Permission | Needed for |
|---|---|
| Microphone | Recording at all. Without it the app records silence |
| Accessibility (macOS) | Inserting the text automatically. Without it the dictation stays on the clipboard |
Each row shows its state: granted, not granted, or not requested yet. Allow access fires the one system prompt while the DeutschlandGPT window is in front, which is much better than having it appear mid-dictation over the app you were dictating into.
macOS often only picks up a newly granted permission after the app restarts. The panel offers Restart now for exactly that. If dictation still does not work after granting, restart before looking for anything else.
Two platform notes. On Windows there is no state to read: the system asks on your first dictation, and a blocked microphone surfaces as a real recording error rather than as a status. So Windows gets the settings shortcut and the explanation, never a claim about a state the app cannot verify. If the microphone is blocked by device management, the row says so and points you at your IT team, because no click in the app can lift that.
Trigger: push to talk or toggle
In Settings, Shortcuts, right under the system-wide hotkeys.
| Mode | Behaviour |
|---|---|
| Push to talk | Records only while you hold the keys, and stops the moment you let go |
| Toggle | One press starts, the next press stops |
Push to talk suits short dictations and never leaves a recording running by accident. Toggle suits long ones, where holding a chord for two minutes is uncomfortable. The Ctrl+Alt+D default was picked partly because it is comfortable to hold.
Insert automatically
On, the result goes straight into the app you were last using. Off, it only goes to the clipboard and you paste it yourself.
Turn it off if you dictate into applications that react badly to simulated keystrokes, or if you would rather look at the text before it lands in a customer-facing email.
Keep in clipboard
Only appears while Insert automatically is on, because without automatic insertion the clipboard is the delivery channel and the text always stays there.
On, the dictation stays in the clipboard after being inserted, so you can paste it again. Off, your previous clipboard contents are restored, which matters if you were in the middle of copying something else.
How long the result stays visible
The transcribed text stays in the dictation overlay for the chosen duration and then disappears on its own: 3, 5, 10 or 20 seconds, or stay open.
Fixed steps rather than a free field, because a free millisecond value invites values that are either unreadably short or leave the overlay parked on top of another app. Pick stay open if you want to read and copy the text yourself every time.
Mode
A mode decides what happens to the raw transcript before it is inserted.
| Mode | Result |
|---|---|
| Raw text | The transcript verbatim, no model call. The fastest path |
| Cleaned | Filler words, repetitions, self-corrections and stutters removed, punctuation and capitalisation fixed. Wording, tone and content untouched |
| Message | A short, casual chat message. No salutation, no sign-off, no bullet lists unless you dictated one |
| A polite, clearly structured email in short paragraphs. No salutation or sign-off is invented if you did not dictate one, and no content is added | |
| Note | Condensed into short bullet points, keeping every fact, name, number and date |
| AI prompt | A precise, complete instruction to an AI model. The intent stays exactly as dictated, no new requirements, and the instruction is not answered |
| Custom | Your own instruction |
Every mode keeps the language you spoke, so dictating in English gives English back.
Raw text is also the fallback: if refinement fails for any reason, you get the plain transcript rather than nothing.
Custom instruction
Only appears while the Custom mode is selected. Whatever you write here is the instruction passed to the model, for example "rewrite as a factual email in English, without greeting or sign-off".
Left empty, the custom mode behaves exactly like Raw text instead of sending an empty instruction to the model.
Vocabulary
Proper nouns and product names are what speech recognition reliably gets wrong. Enter them here, one term per line, and obviously mis-transcribed variants of those words are corrected without the rest of the text being touched.
One per line rather than comma-separated, because that is the only format that survives terms containing commas or spaces.
The vocabulary is passed along with the mode instruction, so it has no effect in Raw text mode: that mode makes no model call at all.
History
Your recent dictations, newest first, with their recordings. Per entry you can:
- Copy the text to the clipboard again.
- Reprocess it with the currently selected mode, which replaces the listed entry. This needs the raw transcript, so an entry without one reports that instead of silently doing nothing.
- Play or download the recording.
- Delete it, on the server and locally.
The history exists for one specific failure: the paste can land in the wrong window, and by the time you notice, the clipboard already holds the next dictation.
Entries are stored on the server, so they survive a reinstall and are visible on your other machines, and they are governed by the retention periods under Data storage. Recent local entries are merged in on top, which covers being offline and the seconds between a dictation finishing and its upload landing.
Clear history deletes every listed dictation and its recording permanently, including on your other devices.
An interrupted dictation, where the app crashed or was quit mid-recording, is recovered and processed at the next start and marked as recovered in the list.