Catch the Translation Failures You Can Actually Prove
Paste the source transcript on one side and the translation on the other, and it compares them segment by segment. It reports what is missing, what was left in the original language, which numbers vanished, and where the subtitle timings stopped lining up. All of it in the browser, so an unreleased script stays on your machine.
This exists because machine translation fails in a specific, boring, checkable way: it quietly drops things. A file goes in with 40 cues and comes back with 31, and nothing in the process tells you. Nobody notices until a viewer does.
Missing segments
Fewer translated cues than source cues, the most common failure.
Lines left untranslated
Segments passed through in the original language.
Numbers that vanished
Prices, dates and percentages that did not survive.
Timing drift
Subtitle cues that no longer start at the same moment.
What a Report Looks Like
This is the shape of the result. The files, lines and numbers are made up for this example.
webinar-en.srt vs webinar-es.srt
Missing segments 3 missing
Source: 40 segments
Translation: 37 segments
The translation has fewer segments than the source. Content was dropped, usually at the end.
Left untranslated 00:02:14
Source: Click Save to keep your changes.
Translation: Click Save to keep your changes.
Identical to the source. Either it was skipped, or it is a proper noun that stays the same.
Numbers dropped 00:05:38
Source: Prices start at 49 dollars a month.
Translation: Los precios empiezan en dólares al mes.
These numbers are in the source but not the translation: 49.
What It Checks, and Why Those Things
Every check here is something you can settle with evidence rather than opinion.
Missing segments
The count of translated cues against the source. On long files it is almost always the end that is gone, and the start reads perfectly.
Lines left untranslated
Text identical to the source. Sometimes right, like a brand name. Often a chunk the translator skipped.
Numbers that disappeared
Figures pulled from each source segment and checked in the translation. A fluent sentence can still lose the number that mattered.
Timing drift
When both files carry timings, the cue starts are compared. A line at the wrong moment is worse than a clumsy one.
These are the exact limits it uses.
| Check | Warning | Problem |
|---|---|---|
| Segment count | More segments than the source | Fewer segments than the source |
| Empty or untranslated | None | A blank translation, or text identical to the source |
| Numbers | A number in the source is missing from the translation | None |
| Length | A line under 35% of the source length, or over 260% of the source length | None |
| Timing | Cue start times that no longer match | None |
Sources, checked 2026-10-12: Translation Checker settings
The score starts at 100, minus 120 points for every problem and 45 for every warning, each divided by the number of cues, counted per segment.
What This Cannot Do, Said Plainly
It does not judge whether the translation is any good.
Fluency, tone, register, whether a joke landed or a term of art was rendered correctly: none of that is checkable by comparing two strings across languages. A segment can pass every check here and still be wrong in a way only a speaker of the language will catch.
So treat this as the first gate, not the last. It catches the mechanical failures cheaply and at scale, which frees a human reviewer to spend their attention on meaning instead of counting cues. I would run this before sending anything to a reviewer, not instead of one.
Use it as the first gate
Run the check
Paste both files and fix every problem it finds.
Glance at warnings
Numbers and length flags need a quick human look.
Then send to a reviewer
Their time goes on meaning, not counting cues.
It proves the translation is complete. Only a speaker of the language can prove it is right.
The length comparison is the softest check for the same reason. Some language pairs genuinely expand or compress, German against English being the obvious one, so a flag there is a prompt to look, not a verdict.
Good at
- Finding dropped and empty segments
- Spotting lines left in the source language
- Catching numbers lost in translation
- Comparing cue timings between SRT or VTT files
Not built for
- Quality: fluency, tone and terminology need a human
- Word errors: the words are meant to differ, so it cannot count them
- Brand names: identical lines are flagged even when they are correct
- Translating: it checks a translation, it does not make one
Transcription Accuracy Is a Different Question
Worth separating two things that sound alike. Checking a transcript against the audio it came from, same language, is transcription accuracy, and the measure for that is word error rate. The word error rate calculator does that job: it aligns two versions of the same language and counts substitutions, deletions, and insertions.
This page is the cross-language version, and it cannot count word errors, because the words are supposed to be different. That is why it measures structure and completeness instead.
Did the transcription hear it right?
Compare a transcript with what was said, word by word. That is word error rate, and the WER calculator measures it.
Did the translation keep all of it?
The words are supposed to differ, so this checks structure and completeness: segments, numbers, untranslated lines and timings.
Before and After the Check
These tools cover the steps around the check.
Translate a video
The video translator handles dubbing and subtitles.
Translate from a link
The YouTube video translator works straight from a YouTube URL.
Make the source transcript
The transcript generator produces it from a recording.
Turn it into subtitles
The text to SRT converter keeps the timings intact.
