Writing check, voice input and translation guide

Three helpers make writing in Runova easier. The writing check underlines spelling, grammar and style problems in text boxes and offers fixes. Voice input turns what you say into text using a speech model that runs on your computer. Translation converts your text into another language, and Prompt Builder uses it to turn a request in any language into English. All three are set up in the AI tab of Settings.

Where to find it

Click Settings in the sidebar and open the AI tab. Below the Providers card you will find Voice input, Writing check and Translation retries. Type grammar, whisper, voice or translate in the Settings search box to jump straight to them. The AI providers card at the top is separate and is not needed for any of these three.

Saving your changes

Writing check saves each change the moment you make it. Voice input and Translation retries work like the rest of the Settings page: a change shows “You have unsaved changes” and you press Save changes (or Ctrl+S, Cmd+S on macOS) to keep it, or Discard to drop it.

Writing check

The writing check uses LanguageTool to find spelling mistakes, grammar errors, punctuation problems and style suggestions. It works in text boxes across the app, including dialogs and the floating desktop widgets.

Turn it on and tune it

In the Writing check card of Settings, AI tab:

Setting	What it does
Check my writing	The master switch. Off means nothing is checked and no text leaves your computer. On by default.
Underline as I type	Checks a field shortly after you stop typing (about a second and a half) and underlines the problems. Off leaves only the right-click check. On by default.
Include style suggestions	Adds LanguageTool’s stricter “picky” level: wordiness, repetition and typography on top of real mistakes. Off by default.
Language	Auto-detect (the default) works out the language of each field. You can also fix it to English (US), English (UK), German, French, Spanish, Portuguese, Italian, Dutch, Russian, Polish, Ukrainian, Arabic or Chinese.
Native language	Turns on false-friend rules, which catch mistakes speakers of that language tend to make. Choose None to leave them off.

LanguageTool has no Persian rules. Persian text falls back to the built-in spellchecker (see below).

The last three controls are greyed out while Check my writing is off.

Where the checks run

Under Where checks run you pick one of two sources:

LanguageTool online: no setup. The text you check is sent to api.languagetool.org. The free service limits a check to the first 20,000 characters.
Local server: offline and unlimited. It needs the LanguageTool download and Java 17 or newer. A local check covers up to the first 60,000 characters.

When a text is longer than the limit, the writing check panel says only the first part was checked.

Set up the local server

Pick Local server to see the setup panel. It shows badges for the server state (Server running, Starting…, Failed, Stopped), the LanguageTool version it found, and the Java version it found.

Click Download to get the LanguageTool desktop zip.
Click Open tools folder and extract the zip into it. A folder named like LanguageTool-6.x inside is fine. The panel prints the exact folder path.
Click Start server. It takes a few seconds while the rules load. Runova stops the server when it quits.
Use Re-check if you changed something and the badges are out of date, and Stop server to stop it by hand.

The Port field (default 8081, allowed range 1024 to 65535) is the port the server listens on. If something is already serving LanguageTool on that port, for example in Docker, Runova uses it as it is.

Start server stays disabled until LanguageTool is in the tools folder and Java is found. If your Java is older than 17, the panel warns you: install a newer JDK or stay on the online check.

Muted rules

When you choose Never flag this rule from the writing menu, the rule shows up as a chip under Muted rules. Click a chip to unmute that rule.

The issue counter and review panel

Larger writing fields (for example the request box in Prompt Builder, the project Build Prompt dialog and Markdown editors) draw a small counter in the bottom right corner. It reads Checking…, No issues, Check failed or a number such as 3 issues. Click it to open the Writing check panel.

The panel header shows the detected language and a Local or Online badge. Each issue lists:

its type (spelling, grammar, punctuation, typography, style) with a colored dot,
the flagged text (click it to jump to it in the field),
the explanation,
up to four suggestion buttons. Click one to apply it.
an X button, “Ignore this one”, to dismiss that issue.

At the bottom, Check again re-runs the check and Fix N mistakes applies the top suggestion for every spelling, grammar and punctuation issue at once. Style suggestions are never applied automatically, so they stay as you wrote them.

Right-click in any text field (including single-line inputs) to open Runova’s own writing menu instead of the system one. It offers:

spelling or grammar suggestions, first in the list,
Ignore (this one issue),
Never flag this rule (adds it to the muted rules),
Add to dictionary for a misspelled word (not on macOS, where the system dictionary cannot be changed from the app),
Next issue (N left) or Go to the other issue, to jump through the remaining problems,
Check the grammar here or Check writing again, which run a check on the spot when nothing has been checked yet.

Below those, every text field’s menu ends with Cut, Copy, Paste and Select all, each showing its usual key. An item is greyed out when it does not apply: Cut and Copy with nothing selected (and always in a password field), Paste with nothing on the clipboard. In a field with nothing to say about the writing, such as an empty one, the menu is just those four.

When there is nothing to report, the menu says “No writing issues here”. Use the arrow keys and Enter or Tab to pick an item, and Escape to close it. If the menu holds only the four edit commands, nothing is highlighted until you press an arrow key, so Enter and Tab carry on as normal. Typing or deleting closes the menu.

Spellcheck

Separately from LanguageTool, every text box has the built-in spellchecker of the app’s browser engine. It draws a red wavy underline under misspelled words. It uses the languages of your operating system first and always adds English (US), with at most three dictionaries active at once. On macOS the system spellchecker is used instead. Right-click a flagged word to see its suggestions in the writing menu.

Fields that hold things like keys and URLs turn spellcheck off on purpose.

Voice input

Voice input lets you dictate a request instead of typing it. The speech model (Whisper) runs on your computer, so your audio is not sent to a speech service.

Where you can dictate

The Prompt Builder page and the Build Prompt dialog of a project have the same two microphones:

Dictate, next to Your request, adds what you say to your request.
Speak prompt, above Generated prompt, is for when you already know exactly what you want to say: speak in English and your words become the prompt as they are, with nothing generated and nothing translated. It always listens for English, whatever language Dictate is set to.

Only one microphone records at a time; the other one waits until you are done.

The buttons are hidden if your system cannot record audio. The pinned Build Prompt desktop widget has no microphone button.

Dictate a request

Click the microphone button. The first time, allow microphone access when asked.
Speak. The button turns red and says Stop with a running clock, such as Stop 0:12.
Click it again. The button shows Transcribing…, then your words are added to the end of whatever is already in the box, separated by a space.

You can throw a recording away instead: press Esc or click the × next to Stop. Nothing is transcribed, and in the dialog, the dialog stays open.

A long recording on a big model can take a while to transcribe. The rest of the app keeps working in the meantime, because the speech model runs in a process of its own. To give up on it, press Esc or click the × next to Transcribing… (for Speak prompt, the Cancel button over the prompt works too). If the speech model ever stops answering, Runova stops it after a generous wait and says the transcription took too long.

If the recording is very short or silent, nothing is sent to the speech model and you see “Didn’t catch anything. Try again a little closer to the mic.” This keeps the model from inventing words for silence.

Dictation language

The chip next to Dictate (EN, FA and so on) shows the language Dictate listens for. Click it to pick another; the choice is saved as your Spoken language setting. The speech model has to be told the language up front, so speak the language the chip shows.

If the chip shows a warning sign, the current model is too small for that language. Persian needs Large v3 Turbo: the smaller models write Persian as a jumble of half-right words. Open the chip and pick Use a model that understands Persian to switch.

The very first use also downloads the speech model, and the button shows the percentage downloaded. The download goes through the proxy set under Settings, Network, if there is one. If it fails partway, the message says so (“Could not download the speech model”); files that finished downloading are kept, so the next try only fetches the ones still missing. The model is downloaded once and then kept on disk, so later dictation works without internet. Long recordings are processed in 30 second pieces. Large v3 Turbo is slower than the others: on a typical laptop processor expect about two seconds of waiting for each second you spoke, and a short extra pause the first time it loads after the app starts.

If the dialog is closed while you are recording, the microphone is stopped and the recording is still transcribed into that project’s saved request.

If something goes wrong a message appears, for example “Microphone access was blocked. Allow it to use voice input.” or “No microphone was found.”

Voice input settings

In the Voice input card (“Local Whisper transcription for Prompt Builder. The model downloads once and stays cached.”):

Setting	Options
Model	Tiny (fastest, about 75 MB), Base (balanced, about 145 MB, the default), Small (about 490 MB), Large v3 Turbo (most accurate, about 760 MB, the one to use for Persian). A bigger model understands you better but is slower and takes more disk space.
Spoken language	English, Persian, Spanish, French, German, Arabic, Chinese, Russian, Hindi or Turkish. There is no auto-detect: the local model cannot tell languages apart, and left to guess it treats everything as English, which turns Persian speech into made-up English. Settings saved as Auto-detect by older versions now use your system language, or English.

Switching the Model downloads the new one the next time you dictate.

Translation

Translation turns text into another language. It needs an internet connection and no account or API key. The text is sent to Google’s free online translation service, so do not translate anything you must keep private.

Where translation is used

Prompt Builder page: the Translate button under “or translate directly” converts your request into the language you choose in the picker next to it (English, Persian, Spanish, French, German, Arabic, Chinese (Simplified), Russian, Hindi or Turkish). Generate Prompt also translates your request to English first, whatever language you typed in, before the prompt is written.
Build Prompt dialog of a project: Translate copies your request into English without generating a prompt.
A project’s Blueprint and the “Describe <project>” dialog (shown when Runova scaffolds a new project’s files) translate your text to English the same way.

Long text is sent in pieces of about 1,200 characters, one after another, and put back together with your line breaks kept. Each piece gives up if the service does not answer within 15 seconds.

Translation retries

If a translation request fails, Runova tries again. In the Translation retries card (“Extra attempts Prompt Builder makes if a translate request fails.”) enter how many extra attempts you want, from 0 to 10. The default is 3. With 0, a failure is reported straight away. There is a short, growing pause between attempts. Press Save changes to keep the number.

If every attempt fails you see “Translation failed. Check your internet connection and try again.”

Tips

If the online writing check feels slow or you do not want text to leave your computer, switch to the Local server and install Java 17 or newer.
Dictating in a language other than English? Set Spoken language explicitly. It is more reliable than auto-detect for short phrases.
Use Tiny for quick commands and Small for long, careful dictation.
If the writing check never underlines anything, check that Check my writing is on and that Underline as I type is on. You can always right-click for a check on demand.