Translation

createTranslator() translates text between languages. By default it uses Helsinki-NLP's Marian opus-mt checkpoints, one model per direction (~105 MB each, 8-bit), resolved from the language pair and downloaded on first use.

import { createTranslator } from 'ngx-transformers';

readonly translator = createTranslator({ from: 'en', to: 'ru' }); // Xenova/opus-mt-en-ru

const russian = await this.translator.translate('The model runs entirely in the browser.');
const german = await this.translator.translate('Good morning.', { to: 'de' }); // Xenova/opus-mt-en-de, its own download

Try it: translation story.

API

• translate(text, { from?, to? }): resolves to the translated string. The pair defaults to the one given at creation; a different pair uses, and loads, another model.
• load(pair?): downloads the model for a pair ahead of time.
• handleFor(pair?): the underlying PipelineHandle for a pair, for apps that keep several warm and want one progress line per model.
• modelFor(pair?): the checkpoint id a pair resolves to.
• dispose(): frees every loaded model.

The translator's status, progress, error, ready and busy signals mirror the handle of the most recent call, so one <ngx-model-progress> covers every pair.

Language codes follow the checkpoint: ISO 639-1 for opus-mt (en, de, ru, fr, es, zh). A pair with no opus-mt model on the Hub fails to load; the Xenova/opus-mt listing shows what exists.

Multilingual models

NLLB and M2M100 serve every pair from one model and take the codes per call as src_lang and tgt_lang. Pin the model and pass the pair on each call:

readonly translator = createTranslator({ model: 'Xenova/nllb-200-distilled-600M' });

await this.translator.translate(text, { from: 'eng_Latn', to: 'tgk_Cyrl' }); // FLORES-200 codes
await this.translator.translate(text, { from: 'eng_Latn', to: 'deu_Latn' }); // same model, no new download

The library recognises NLLB, M2M100, mBART and MADLAD checkpoints by name and forwards the codes only to those; opus-mt models translate one fixed direction and receive none.

NLLB-200 distilled 600M is roughly 600 MB at q8, so it suits apps that need many pairs rather than one.

Overriding a pair

Point a pair at another checkpoint for the whole app:

provideTransformers({ translationModels: { 'en-ru': 'my-org/en-ru-tiny' } });

Tips

• Sentences translate better than paragraphs. chunklet's chunkSentences(text, { maxTokens: 64 }) splits on real sentence boundaries (abbreviations included) into pieces you can translate in sequence and join back.
• The first call per pair pays the download; call load({ to: 'de' }) early for pairs you know you will need.