Skip to main content

Translator

The Translator is a cartridge: an app installed from MnemoHub. It translates typed text, documents and web pages. It keeps a history and a vocabulary of the words you learn. You pin it to the rail like any other cartridge, see the menu bar.

At the 2.2.0 release, the Translator was at version 1.3. That version has three tabs: a dashboard with a quick translation and the history, a files queue, and a word repertoire.

Arrives in the next build

Translator 2.0 and its far view arrive in the next build. The rest of this page describes them.

The bar on top​

One bar stays on screen whatever the tab.

  • The source language. Detect is the default. Once a text is typed, the menu names the language it found.
  • My languages: one chip per language you use. The green chip is the target. A click on another chip sets the target. A click on the target chip swaps source and target, once the source language is known.
  • Other: a menu with every other language.
  • The engine pill. It opens the list of engines: Auto, which follows the app's routing, Cloud with the cloud model's name, and the Local models installed on this computer.
  • Free memory: unloads the engines left idle, such as Whisper, the voice or the embedder. The status line names what was unloaded.
  • Settings: a drawer on the right.
  • A dot: green when the cartridge is connected to Mnemosyne OS.

With Detect as source, the target follows your languages:

  • A text in one language of My usual pair goes into the other one.
  • A text in another of your languages goes into your first language that differs from it.
  • A text in a language outside your list goes into your first language.

Pick a target by hand and it stays until you clear the text.

The engine is chosen for the Translator alone. It never changes the chat's model. A local model loaded here does take the chat's place in memory. The chat loads its own model again for its next question.

The six tabs​

Translate, Documents, Web page, Compare, Vocabulary and History, in that order. A tab once opened keeps its state while you look at another one.

Translate​

Type or paste a text, then press Ctrl+Enter or Translate. A .md or .txt file can be opened here too. PDF and DOCX go in Documents.

A long text goes piece by piece, with a progress line and Stop. A stop keeps what was translated so far. Under the result:

  • Words collects the words worth learning into your vocabulary.
  • Discuss in chat hands the translation to the chat.
  • Export .md saves it as a Markdown file.

Documents​

Add a file queues PDF, DOCX, EPUB, TXT and MD files. Translate all runs them one after the other: read, then translate piece by piece. For scanned pages, tick the OCR box in the bar. Reading is slower then. A finished file can go to the chat or be exported.

Web page​

Paste an https:// address and press Read. The page is read on this computer. The address goes to no other service. Translate the page then builds a report with the original and the translation. The report has three exits:

  • Discuss in chat hands it to the chat.
  • Keep stores it in the Translator's own vault.
  • Export .md saves it as a file.

The Translator's vault is a sandbox. You make what it holds permanent from the vault manager.

Compare​

A bench for engines. Give it one text and two to six engines in order, then press Run. The engines run one at a time: load, translate, unload, next. The table shows the load time, the translation time, the speed in characters per second and the length.

Agreement counts the words two translations share. It measures agreement, and two engines can agree on a mistake. Differences shows two translations word by word. Scenarios are saved with Save, and the results leave as Report .md or in the chat.

While a comparison runs, the other tabs wait for it.

Vocabulary​

The words collected with Words, searchable and filtered by language pair. Review opens flashcards for the words due now. With none due, it takes them all, the oldest first. Each answer moves the card to a box, and the box sets when it comes back.

History​

Every translation kept: typed text, files and web pages. Figures sit on top: translations, documents, words and languages. Open puts a translation back in the Translate tab. Delete all asks first.

Settings​

The drawer holds:

  • My languages, in order, with Move up and Add.
  • My usual pair: the two languages that translate into each other.
  • Engine.
  • Free memory before loading a local model.
  • Piece size (characters): 3,500 by default, between 500 and 12,000.

What it keeps​

The history and the vocabulary live in the cartridge. A copy of each translation and each word also goes to the Translator's sandbox vault. Its tile counts Translations and Words.

From across the board​

Zoom the canvas out and the window becomes a card. It shows Translate, your language chips with the target highlighted, the last result and the engine.

A click opens the quick translation in a panel at normal size, whatever the zoom. Type a text and press Enter. The target follows your languages like in the window. Other offers any language. Copy takes the result, and Open the Translator brings the camera to the window.

The quick translation takes up to 4,000 characters. Over that, it asks you to open the Translator, which translates piece by piece.

Night translation, for developers​

Night translation is a separate window. It translates the locale files of an app you build. Open it from MnemoHub, tab Dev, section Your tools.

A local model translates the strings copied from English while you sleep. Nothing is written into the locale files until you apply. The night never falls back to the cloud: with no local model installed, it does not run.

  • Locales folder: the folder that holds en.json. Then Save.
  • Run every night: the pass starts in a night window, after a stretch of idle time. The window shows the hours.
  • A trial button runs a small batch first. Run everything now runs the whole pass at once. Stop ends a pass.

The morning report gives, per language, the strings proposed, to look at, refused and waiting. Two lists follow:

  • Random check draws a few proposals the checks let through.
  • To look at lists the ones the checks flagged: same as English, unusual length, different line breaks, or a product name that may be missing.

Each proposal takes Accept, Fix, Reject, or Keep as is when it matches English. The apply button counts the accepted strings. It asks once more, then Write them writes them into the locale files. Nothing is committed.