Polyglot for NVDA

Polyglot is an NVDA global add-on focused on fast, flexible multilingual translation. It can translate selected text, clipboard content, and the last text spoken by NVDA, and it can also intercept spoken content for live auto-translation.

The add-on is built around a dynamic engine architecture. Translation engines declare their own capabilities and configuration schema, and the settings UI is generated from that schema at runtime. That keeps the core plugin small while making it straightforward to add new services.

What It Does

• Translates selected text, clipboard text, and the last spoken NVDA utterance.
• Provides a translation command layer on NVDA+Alt+Z; press H in the layer to show command layer help.
• Supports live auto-translation of spoken NVDA content.
• Includes a smart speech filter to avoid translating roles, states, and formatting noise.
• Persists a translation cache to reduce repeated requests.
• Can copy manual translation results to the clipboard automatically.
• Lets you switch engines and languages without leaving the keyboard.
• Exposes a dedicated interactive translation dialog for longer or iterative translation work.
• Provides local English word definitions when NVDA is running in Chinese.

Installation

The preferred installation path is the NVDA Add-on Store. You can also install manually:

1. Download the latest .nvda-addon package from the Releases page.
2. Open the downloaded file.
3. Confirm installation in NVDA.
4. Restart NVDA when prompted.

Quick Start

1. Open NVDA menu -> Preferences -> Settings -> Polyglot.
2. Choose a translation engine and make sure it is enabled.
3. Configure any required credentials for that engine.
4. Set source and target languages.
5. Optionally enable clipboard copy and the smart speech filter.
6. Press NVDA+Alt+Z, then use one of the command-layer keys below. Press H in the layer to show command layer help.

Command Layer

Press NVDA+Alt+Z to enter the command layer. A short beep confirms that the layer is active. Press H in the layer to show command layer help. Most commands execute once and exit the layer. Language and engine switching commands stay inside the layer so you can continue cycling. Engine switching cycles through enabled engines only.

Interactive Translation Dialog

The interactive dialog is designed for longer text and iterative translation work.

• Open it from the command layer with I.
• Select an enabled engine, source language, and target language without leaving the dialog.
• Disabled engines remain configurable in settings, but are not listed in this dialog.
• For LLM-style engines, adjust model and prompt template directly in the dialog.
• Press Ctrl+Enter in the source text box to translate.
• Copy the result or clear both panes without reopening the window.

Local English Word Definitions for Chinese NVDA

When NVDA’s interface language is Chinese, including Simplified and Traditional Chinese, Polyglot adds an offline English-to-Chinese dictionary to the command that reports the word at the review cursor:

1. Press numpad5 once in the desktop keyboard layout to read the word at the review cursor.
2. Press it twice to hear the word spelled out.
3. Press it three times to hear the Chinese definition when the complete word is found locally.

Add-ons that call NVDA’s speech.spellTextInfo receive the same behavior. If a line contains only one English word, pressing the current-line command three times also reports its definition. Lookup handles one layer of common leading or trailing sentence punctuation, safe spelling variants, and common plurals, tenses, participles, comparatives, and superlatives. Exact entries take priority. When a spelling can refer to several entries, Polyglot announces up to three candidates with their first senses. If a two- or three-character uppercase word only matches a lowercase entry, Polyglot reports the possible lowercase meaning and warns that it may instead be an abbreviation. Words that look valid but are not included are reported as missing. Single characters, multiword lines, and unsupported text keep NVDA’s original character descriptions. Lookup is fully local and never sends the word to a translation service.

The Translate Selection, Translate Clipboard, and Translate Last Spoken commands also use the local dictionary for complete English words in English/auto-detect → Simplified or Traditional Chinese requests, and for English-word input in Chinese → English requests. Other content continues through the selected translation engine.

Translation commands and text review have separate local-dictionary options. Both are enabled by default; disabling either restores that command’s original behavior.

Current dictionary size:

• 122,370 English entries, including about 118,460 complete headwords that can be queried directly.
• 34,859 validated inflected forms.
• About 154,800 complete word forms when case and common spelling variants are included.
• The bundled dictionary is about 6.8 MB and requires no separate download.

Settings Guide

Common Settings

• Copy manual translation results to clipboard: Copies manual translation output after a successful request.
• Prefer the local English-Chinese dictionary for selected text, clipboard text, and last spoken text: Controls local lookup for those three translation commands.
• Use the local English-Chinese dictionary in text review: Controls local definitions for NVDA text review.
• Enable smart speech filter: When translating spoken NVDA output, skips non-content speech such as roles, states, location, and formatting details where possible.
• Clear Cache: Clears the persistent translation cache and shows the current item count in the button label.

Shared Engine Settings

Most engines inherit a common set of settings:

• Enable this engine: controls whether the engine is available for translation requests, command-layer engine switching, and the interactive dialog. Disabled engines remain visible and configurable in settings.
• Source language and Target language
• Proxy mode: use system proxy settings or disable proxy usage
• Request timeout

If an engine reports detected source language, Polyglot also exposes:

• Auto-swap if detected source matches target: useful when the source language is set to auto-detect
• Swap to language: the alternative target used during auto-swap

Auto-Translation Behavior

• Auto-translation acts on spoken NVDA content captured by the speech pipeline.
• The add-on suppresses its own spoken messages to avoid translation loops.
• If auto-translation fails three times in a row, it is turned off automatically.
• The smart speech filter mainly affects spoken-content translation, not standard manual text translation.

LLM and Polyglot-Specific Options

Some engines expose additional controls:

• Ollama 1 and Ollama 2 provide two separate saved profiles for different local or remote Ollama setups.
• OpenRouter exposes API URL, API key, model preset, custom model name, prompt template, and custom prompts.
• Ollama engines expose API URL, model name, optional API key, prompt template, and custom prompts.
• Google Translate (Polyglot) exposes a configurable endpoint URL and API key field.
• Google Translate (key-free) offers an optional mirror-server toggle.

Chrome AI Offline Translation

Polyglot can use Chrome’s built-in Translator API for offline translation. Translation is handled by an isolated local Chrome instance, so the text is not sent to a third-party translation service.

Requirements

• Google Chrome must be installed.
• Chrome 138 or later is recommended.
• The first use of a language direction requires the local translation model to be prepared. Polyglot can download the model through its model manager, or you can let Chrome download it.

How To Use

Select Chrome AI (Offline) in Polyglot settings, then choose the source and target languages. Chrome AI requires an explicit source language; Auto-detect is not available for this engine, so Polyglot can check the required model before starting Chrome.

On first use, if the required model is not installed, Polyglot asks how to proceed. Choose Yes to download and install the model with Polyglot’s model manager; use this if Chrome’s model download service is slow, blocked, or unreliable on your network. Choose No to let Chrome download the model. Choose Cancel to cancel the current translation. After the model is ready, translation continues automatically.

Network And Models

Translation runs locally. Models can be installed by Polyglot’s model manager or downloaded by Chrome. If Chrome’s model download service is slow or unavailable on your network, choose Yes in the prompt, or open the Polyglot ChromeAI model manager from NVDA’s Tools menu to install or remove offline models in advance.

Privacy And Data

Polyglot uses a separate Chrome data directory for Chrome AI, so it does not affect your regular Chrome profile. Models, cache data, and runtime data are kept to avoid repeated downloads.

The default location is:

%LOCALAPPDATA%\Polyglot\ChromeAI

If the LOCALAPPDATA environment variable is not available, Polyglot falls back to the polyglot_chrome_ai directory under the NVDA configuration directory.

When NVDA exits, Polyglot closes the Chrome instance it started.

Limitations

• Supported languages and language pairs are determined by Chrome’s Translator API.
• Chrome AI requires an explicit source language; Auto-detect is not available for this engine.
• First use requires the model to be prepared; model downloads may be affected by network conditions.
• If the Translator API is unavailable, update Chrome or make sure the related Chrome feature is enabled.

Engine Overview

The repository currently includes the following engines:

Contributing

Contributions are welcome across code, documentation, localization, testing, and engine integrations.

• Issues: GitHub Issues
• Releases: GitHub Releases

When adding a new engine:

1. Create a module under addon/globalPlugins/polyglot/services/engines/.
2. Implement TranslationEngine or, for HTTP engines, extend BaseHttpEngine.
3. Return a config spec from getConfigSpec() if the engine needs settings.
4. Use supported control types from views/factory.py: choice, text, password, checkbox, and spinctrl.
5. Verify the engine appears correctly in the dynamic settings panel and, when enabled, command-layer switching and the interactive dialog.

License

Copyright (C) 2025-2026 cary-rowen.

This project is licensed under the GNU General Public License version 3 or later (GPL-3.0-or-later). See the repository’s COPYING.txt.