Machine translation engines

A machine translation engine takes a sentence in one language and produces a sentence in another. Most engines today are neural models trained on large collections of translated text, and a growing number also use large language models. You reach them in two ways: through a web or desktop app where you paste text or upload a file, or through an API that your own software, translation management system or website plugin calls in the background.

The engines in this category fall into three groups. Cloud providers such as Google Cloud Translation, Azure AI Translator and Amazon Translate sell translation as one service among many, billed through the same account as your hosting. Specialist vendors such as DeepL, ModernMT, SYSTRAN, Language Weaver and Lingvanex focus on translation and often add features aimed at businesses, such as on-premises deployment or engines that adapt to your corrections. Open source projects such as LibreTranslate and Apertium let you run translation on your own servers, which matters when text cannot leave your network.

Quality varies by language pair and by type of text far more than by brand. An engine that reads well for English to German marketing copy can stumble on Japanese interface strings or Finnish legal text. Short strings with no surrounding sentence, such as a button that says "Save" or "Charge", are hard for every engine because the right word depends on context the engine never sees. The only reliable comparison is a test with your own content: pick a few hundred real sentences, run them through two or three engines, and have native speakers mark the errors that would matter to a reader.

Customization is the next thing to check. Glossaries force specific translations for product names and key terms. Custom models, offered by Google, Microsoft, Amazon and others in different forms, are trained on your own translation memories so the engine learns your vocabulary and tone. Adaptive engines such as ModernMT update from each correction a linguist makes. Ask which of these work for your language pairs, since support is often uneven.