Sinch Translation Quality
Undertow Sinch Sinch
Monthly quality report

Sinch translation quality

What the AI translation is getting right on its own, what the language team is still having to change, and what we have altered in the setup to close the gap.

Last updated

The headline

How much the AI output needed changing

Of everything the AI translated this cycle, this is the share a linguist had to genuinely change. As the glossary, the AI instructions and the style guide improve, this share should fall. It is shown per language, because the four behave differently and an average would hide that.

The record so far

Every cycle since the programme started

Share of segments needing a real edit, by cycle

Hover any cell for the exact figure.

Less changed More changed

Where the errors are

What kind of errors they are

Real errors by category

Ranked by how often they occur. Bars are split by language.

Show

What we changed

What we changed, and whether it works

Show
ChangeLanguageTypeLive since Status

Rules that cannot be checked this way

Some changes are matters of judgement rather than a right answer we can search for. These are assessed in the monthly review instead.

For the Sinch team

Where the English source can help

ItemTypeExampleRaisedOpen

Coverage

What these numbers are based on

How much material sits behind this cycle, so you can judge how much weight the figures above carry.

AI and spend Kept apart from the quality tabs on purpose: quality is judged on its own, never against cost.

AI models and spend

What the AI costs, and what it buys

AI spend per month

Split between the AI Proofreader and the translation step (including suggestions in the editor). Hover a bar for the figures.

AI Proofreader Translation and suggestions

Price per 1,000 words, and the same words translated by people

Fixed a real error Preference Made it worse Not sure