windytranslate

Find, license-check and run Windstorm Labs' open translation and speech models: 1,660 translation pair models, 3 multilingual models and 66 speech-to-text models, each with its measured score, licence and attribution. Catalogue: windytranslate.com.

pip install windytranslate            # catalogue client: standard library only
pip install "windytranslate[local]"   # + local translation with transformers

import windytranslate as wt

m = wt.find("en", "de")       # best measured English -> German model
print(m.repo, m.licence, m.chrf, m.page)
print(m.notice)               # attribution text to ship with your product

print(wt.translate("Where is the train station? The museum opens at ten.", "en", "de"))

windytranslate find en sw --all
windytranslate find en fr --notice
windytranslate translate en fr "Where is the train station?"

• Scores are a FLORES-200 screening (48 sentences per pair), published for every scored pair including weak ones. A model with no score is "not yet scored", never "good". Method: windytranslate.com/evaluation.
• Licences belong to each model (usually Apache-2.0 or CC-BY-4.0, from OPUS-MT by Helsinki-NLP). CC-BY requires attribution; m.notice gives it. m.licence_clear is False when the catalogue flags the licence (e.g. under review). See windytranslate.com/licensing.
• Text is translated sentence by sentence: small pair models drop or invent sentences when a paragraph is one input.
• Everything runs on your machine after the model download. No key, no account.

This package's code is Apache-2.0. Hosted API: early access, free in beta, windytranslate.com/api. Examples and a Docker self-host server: windytranslate-examples.