Mozilla has released version 0.4 of the Firefox extension Firefox Translations (formerly developed under the name Bergamot Translate), implementing a self-sufficient machine translation system that operates on the browser side without relying on external services. The translation engine utilizes bergamot-translator, developed within the Bergamot initiative by Mozilla developers in collaboration with researchers from several universities in the UK, Estonia, and the Czech Republic, with financial support from the European Union. The code is distributed under the MPL-2.0 license.
The engine is written in C++ and compiles to an intermediate WebAssembly binary using the Emscripten compiler. It acts as a wrapper over the Marian machine translation framework, which employs recurrent neural networks (RNN) and transformer-based language models. GPU can be used to accelerate training and translation. Marian is used to power the Microsoft Translator service and is primarily developed by engineers at Microsoft together with researchers from the University of Edinburgh and the University of Poznań.
Firefox Translations supports translation from Estonian and Spanish to English and vice versa, as well as from English to German. The translation speed is 500-600 words per minute. It includes support for prioritizing the translation of text visible in the browser window. The new version provides the ability to automatically download models upon the first translation attempt. Model files take about 15 MB for each language. This automatic downloading leads to a slight delay before the first translation, but drastically reduces the size of the extension itself (3.6 MB instead of 124 MB).
The new version also significantly speeds up model loading into memory — previously, loading a model took 10-30 seconds, but now models load almost instantly. If translating a page takes more than 3 seconds, a progress indicator is provided in the interface. Translation occurs sequentially from top to bottom, starting with the visible area. Translated parts are shown as they become ready, while untranslated parts remain in the original language.
Telemetry transmission has been enabled, including data on user interactions with the extension's interface (for example, about button clicks for translation or disabling translations for specific sites), information about operation execution time, and technical details about the system (CPU, memory).
The extension can currently only be installed in Firefox nightly builds with the digital signature verification for extensions turned off ('xpinstall.signatures.dev-root=true' and 'xpinstall.signatures.required=false' in about:config). After installing the extension, Firefox will start displaying a panel suggesting to translate pages whose language differs from the browser's language and is supported by the extension. There is an option to disable further panel output for a specific language or site.
It is worth noting that Firefox already has a built-in mechanism for translating pages, but it relies on the use of external cloud services (Google, Yandex, and Bing are supported) and is not enabled by default (to enable it in about:config, the 'browser.translation' settings should be changed). This translation mechanism also supports automatic language detection when opening a page in an unknown language and displays a special indicator with an offer to translate the page. The new extension uses the same interface for user interaction, but instead of calling external services, it initiates a built-in handler that processes data on the user's system.
Source: opennet.ru
