NoaLingua for Chrome
Add to Chrome — Free Add to Chrome

Answer

How NoaLingua uses Chrome’s on-device translator

First in a chain of four, and why it is first

Chrome 138 added a translation model that runs on your own machine. NoaLingua tries it first for your language pair, because it is instant, works offline, and translates a sentence the extension assembled itself — so it has no caption-boundary errors.

Short answer On-device first, because it is instant, offline and free of boundary errors. Three named fallbacks after it.

The detail behind the NoaLingua answer

The order, and why it is that order

Four sources in a fixed sequence: Chrome’s on-device model when it is genuinely ready for your pair, then YouTube’s own subtitle translation, then MyMemory if you explicitly chose cloud translation, then YouTube again as a last safety net.

On-device is first because it is instant, offline and translates a sentence NoaLingua assembled itself. YouTube’s translation uses a request the player makes anyway, so no new third party sees the text. MyMemory is the only source that sends sentence text off your device, which is why it sits behind its own checkbox and never appears in the fallback chain.

What "genuinely ready" means, and why it matters

Chrome does not have a model for every pair, and where it does the model may still be downloading. Treating "the API exists" as "the model is ready" is how an extension produces a blank line instead of a translation.

It is also why the minimum is Chrome 138: that is the first release carrying the on-device Translator API the extension depends on. There is no Firefox or Safari build, because there is no equivalent to depend on.

More questions people ask NoaLingua

Do I need to configure the on-device model?

No. If Chrome has a model ready for your pair it is used automatically. If it does not, the chain falls through to the next source without you having to do anything.

Does anything leave my machine when on-device translation is used?

No. That is the point of it being first. Sentence text only leaves the device if you switch translation to Cloud yourself.

Which pairs does Chrome not have a model for?

Chrome does not publish that per language, so the coverage matrix marks it unverified rather than guessing. English to Serbian is one confirmed gap, which is why YouTube’s own caption translation matters as a safety net.