Skip to content

Machine Translation

Summary

Machine Translation (MT) is the automated translation of text from one language to another using computational models, without human intervention. In the open-source ecosystem, LibreTranslate powered by Argos Translate represents the leading self-hosted option.

Types of Machine Translation

Type Description Example
Rule-based (RBMT) Linguistic rules and dictionaries Early MT systems
Statistical (SMT) Statistical models from parallel corpora Moses, early Google Translate
Neural (NMT) Deep learning models (sequence-to-sequence) Modern Google Translate, Argos Translate
LLM-based Large language models with translation capability GPT-4, Claude

LibreTranslate Architecture

Flask API server wrapping Argos Translate (neural MT engine):

  • Translation engine: Argos Translate — open-source neural MT using OpenNMT
  • Language detection: lexilang (short text) + langdetect (long text)
  • File translation: argostranslatefiles for format-aware translation
  • Models: Downloaded on first run or baked into Docker image
  • No proprietary dependencies — unlike Google/Azure APIs

Integration with CAT Tools

CAT tools integrate MT engines as additional suggestion sources alongside translation memory matches:

  • omegaT — MT via plugins (Google Translate, DeepL, Azure)
  • weblate — "Machinery" module with multiple MT providers
  • tolgee — Built-in machine translation providers
  • poedit — Limited MT support

Self-Hosted vs Cloud MT

Self-Hosted (LibreTranslate) Cloud (Google, DeepL, Azure)
Cost Free (compute cost) Per-character pricing
Privacy Full data control Data sent to third party
Offline Yes No
Quality Good (improving) Best-in-class
Languages Limited (community models) 100+ languages
Setup Docker or Python install API key

Argos Translate

The engine behind LibreTranslate: - Based on OpenNMT (neural machine translation framework) - Language packages installed separately - Community-contributed translation models - Can train custom models with Argos Train

See Also