Review translated subtitles before publishing

Check meaning before polishing the wording

Review translated subtitles against the original speech and a checked source-language track, then watch the target subtitles with the final video. First resolve changed meaning, missing information, names and quantities. Only then polish phrasing and line breaks. A sentence can be grammatical, natural and completely wrong about what the speaker said.

The reviewer needs to understand both languages and the subject. If you cannot judge the target language, arrange that review before treating the translation as ready. Reading an automatic translation back into the source language can suggest questions, but it is not independent approval: another translation can hide or introduce an error. Use media and subtitle text you own or have permission to translate and distribute.

Lock the video edit and check the source track

Keep the original video, source subtitles and generated translation as separate files. Give the working copies a shared edit identifier, such as gallery-tour-v3, and a language suffix. Confirm that the subtitles belong to that exact video cut. A translation made from an older edit can be beautifully written while describing a scene that is no longer present.

Check the source transcript against the audio before resolving translation disputes. Listen again to unclear names, negatives, quantities and interrupted sentences. W3C’s caption guidance explains that automatically generated text needs accuracy review and that a missing negative can reverse the message. Translation will not reliably repair a faulty source transcript.

When the audio is genuinely unclear, record the uncertainty and ask the content owner or speaker when possible. Do not fill the gap with a plausible sentence. If the source changes during review, identify every affected target cue and review it again before export.

Separate accuracy, terminology and fluency

The MQM error typology distinguishes accuracy from terminology and linguistic conventions. That distinction makes a practical review easier: first compare what each passage means, then check agreed terms, and then improve the target language. This guide uses that distinction as a reading aid; it does not calculate an MQM score or certify a translation.

For each cue, ask who does what, to whom, when, and under which condition. Check negatives, comparisons, exceptions and degrees of certainty. “May open” must not become “will open”; “fewer than ten” must not become “ten”; a question must not turn into an instruction. Look at neighboring cues because a condition or pronoun may depend on the preceding sentence.

Also check for additions and omissions. A translator may need different word order or more words, so word-for-word matching is not the goal. The target should preserve the message, tone and relevant uncertainty without inserting explanations the speaker never supplied.

A small example: one missing word changes the plan

This fictional English-to-Dutch example is original editorial sample text. The intended message is that the gallery opens on Saturday instead of Friday. The flawed Dutch draft below is readable but removes the exclusion. The correction restores it; neither version is output from a Polyglot accuracy test.

In an actual timed file, make the text correction against the same cue identifier and timestamps, then read the preceding and following cues. If the source actually said “might open,” this correction would still be incomplete: the uncertainty would also need to survive. Review what is in your recording, not merely the pattern in this example.

Source, cue 24:
We open on Saturday,
not Friday.

Flawed Dutch draft:
We gaan zaterdag open.

Reviewed Dutch wording:
We gaan zaterdag open,
niet vrijdag.

Keep a short reference for names, terms and quantities

Create a manual reference sheet containing recurring source terms, approved target wording, names that stay unchanged, and any context needed to choose between meanings. Include the language variety and preferred level of formality. Netflix’s own delivery guidance requires consistency references for its projects; a small independent production can use a simpler sheet agreed with its reviewer. Those Netflix rules apply to Netflix delivery, not every subtitle file.

Check each occurrence in context rather than replacing every matching word automatically. A product name may remain unchanged while the ordinary word beside it should be translated. The same source word can have different meanings in different scenes. This reference sheet is a human review aid; this workflow does not imply a Polyglot glossary-import feature.

Compare quantities as meaning, not just matching digits. Check decimal separators, units, ranges, dates, times and currency labels. Do not silently change a value or convert a unit to make the text look familiar. If localization requires a conversion, have the content owner approve the equivalent and verify it separately. A numerically tidy sentence can still communicate the wrong amount.

Review idioms with the scene, not in isolation

Play the full sentence and surrounding scene when an expression seems odd. Identify whether it is literal, humorous, sarcastic or figurative, and ask what the audience needs to understand at that moment. Choose target wording that carries that function without importing an unrelated cultural reference or changing the speaker’s attitude.

Keep deliberate ambiguity when the source leaves something unresolved. Do not reveal an unnamed person or future event just because later scenes make the answer clear. On-screen writing can clarify a term, but it is not permission to silently rewrite contradictory speech. Record a conflict between spoken and visible information for the content owner to resolve.

Track decisions by cue, then preserve the structure

For each unresolved issue, keep a small manual log: cue ID or timestamp, source wording, target draft, reason for concern, approved correction, and review status. Link the log to the video edit and language. Resolve repeated terminology issues across the track, and keep the earlier export so a disputed change can be traced. This log is a suggested document, not an automatic report from Polyglot.

Polyglot validates the cue count, identifiers, order and nonempty text of translation responses, and retains source timing. These structural checks help prevent a malformed translation from becoming a misaligned track. They do not establish that a negative, name or idiom was translated correctly. Completed translations still need a meaning review.

Source edits can mark affected translations for review without silently replacing completed target text. Recheck those passages and any neighboring context they depend on. After a text-only correction, confirm that no cue was lost, duplicated, reordered or unintentionally retimed. Keep intentional changes explicit.

Approve the delivered track during playback

Read the target track at playback speed after the meaning review. Natural phrasing must also fit the available screen time. If it is too dense, ask the reviewer for a shorter equivalent that preserves the message; do not remove a condition simply to reduce length. The linked line-break guide covers readability in more detail.

Export a separate SRT or VTT file in the format the destination accepts, and test it with the matching final video in that destination. Confirm the language label, accented characters, line breaks, timing and first and last cues, then review the complete playback. A correct local file is not proof that an uploaded player is showing that same revision.

If the deliverable is intended to serve viewers who cannot hear the audio, review meaningful non-speech sounds and speaker information as well. Translating dialogue alone does not complete that captioning task. Hold publication when a meaning-changing issue is unresolved; record the approved edit, language and subtitle filename when the track is ready.

Keep exploring.