Subtitle a recording that switches languages

Keep the spoken source before choosing a translation

When a recording switches languages, first establish what was actually said in each language. Preserve that checked multilingual source separately from any audience translation. Then decide how each delivered track will represent the changes. A smooth English sentence may be a good translation of Spanish speech while still being the wrong text for a source transcript.

This guide covers recorded media you own or have permission to transcribe, translate and share. It uses an original English–Spanish example and a manual review sheet. It does not assume that Polyglot automatically detects every language switch or reliably transcribes or translates mixed-language audio. Arrange reviewers who understand the languages before approving the words.

Define two different files, even when they share times

A source reference records the spoken wording in its actual languages. An audience track provides the wording the intended viewer needs. An English audience track might translate Spanish passages while preserving English speech. A Spanish audience track would make a different set of decisions. Keep those files explicitly named so a translated passage cannot quietly become the supposed original.

This differs from bilingual display, where the original and a translation appear together. A single line containing two spoken languages is not automatically a bilingual comparison pair. Likewise, a complete multilingual source is not a sparse forced-subtitle track. The linked guides cover those separate delivery choices.

Write down the media revision, selected soundtrack, source-reference purpose and intended audience language. If the source reference is only for reviewers, say so. It may need further sound, speaker and presentation work before it is suitable as public accessibility captions.

Listen for changes that the draft may have hidden

Play the recording without relying on the draft to tell you which language you are hearing. Mark each confirmed change and the return to the previous language. Include short phrases inside longer sentences, quoted speech and off-screen voices. A fluent reviewer should establish the actual language; a name, accent or location is not enough to infer it.

Compare those spans with the draft. Look for missing words, a phrase rewritten into the surrounding language, phonetic approximations and duplicated text around the transition. W3C’s transcription guidance treats automatically generated text as material to correct against the audio. A draft that reads naturally still needs that comparison.

Keep a record outside the captions with the media interval, actual language, source wording, proposed audience rendering and review status. If a language or phrase is unresolved, leave it pending. Do not label speech unintelligible simply because the current reviewer does not understand it.

Example: a correct translation can be the wrong source

Imagine one presenter saying the following three phrases in a permission-cleared workshop film. The middle phrase is Spanish; the surrounding phrases are English. All wording, cue numbers and times below are invented editorial examples, not output from an app test.

Suppose the draft replaces “No lo abras todavía” with “Do not open it yet.” That preserves the basic instruction in English, but removes evidence of what language and words the presenter used. Restore the verified Spanish in the source reference. Keep the English rendering in its own audience track. Do not replace the source just because a reader finds the translation easier.

The example keeps the same cue boundaries to make the distinction visible. It does not suggest that every language change starts a new cue, that all translations fit the same duration, or that a source reference is ready for public caption delivery.

Source reference
41 · 00:12–00:14
Leave the box here.

42 · 00:14–00:17
No lo abras todavía.

43 · 00:17–00:20
We will check it together.

English audience rendering
41 · Leave the box here.
42 · Do not open it yet.
43 · We will check it together.

Make the language change available to the viewer

Next decide whether the audience needs to know that the language changed, as well as understand the words. Follow the recipient’s delivery guide and the production’s intended meaning. Netflix’s U.S. English SDH guidance, for example, distinguishes translated speech from speech intentionally left untranslated and requires identification at language changes. Those are Netflix-specific conventions, not universal instructions for every SRT.

For an English accessibility track following that convention, the example’s translated middle cue could begin “[in Spanish]”. The following English cue would need the return identified as required by that same convention. Check both transitions. Otherwise, someone reading captions can be left believing that subsequent English dialogue is still spoken in Spanish.

Dialogue intentionally withheld from the audience is a different decision. Ask the content owner whether it is meant to be understood before revealing a translation. Keep any verified source wording needed for private review separate from the approved public rendering. A label describing speech in another language does not itself translate that speech.

A borrowed word or proper name is not necessarily a language change. Netflix treats foreign words inside a mostly English line separately from full switches, and has specific rules for familiar loan words and names. Use your destination’s convention consistently; do not add a language label mechanically to every unfamiliar word.

Repair the track without inventing a multilingual workflow

Polyglot can import SRT and UTF-8 WebVTT, let you review and edit subtitle text locally, and export separate timed files. Use those established capabilities for a checked source or completed audience track. Verify accents and punctuation after import and export. The app’s recognition language list does not establish automatic switching accuracy within one recording.

If the draft omitted a complete phrase, replacing another cue’s text is insufficient. Author and time the missing cue in a suitable caption editor against the original recording. Keep every interval on that recording’s clock. This article does not rely on automatic language segmentation, cue insertion or combining independently transcribed fragments.

Have the relevant language reviewer approve each source span before approving its target rendering. If a mixed-language input is unsupported or its output cannot be judged, use a qualified human translation workflow and review the resulting track. Do not treat selecting a single source language, repeated generation or agreement between AI drafts as proof that all spoken languages were handled.

Keep the last approved exports. A source correction can affect every audience version, including a version where the line previously looked correct. Recheck those passages and their neighbors without silently replacing completed translations or unsaved edits.

Close each transition in the delivered track

Use a short acceptance note for each switch. Record the media revision and interval, confirmed source words, intended audience rendering, language-identification decision and reviewer approval. Leave the approval field blank until the relevant person has checked it. The note is a suggested manual artifact, not a Polyglot report.

Play into the first changed-language cue from the preceding passage, continue through the return, and repeat at ordinary speed. Check the words, timing, any identifiers and the following cue. Then inspect the exact exported file in the destination player, including its language menu and narrow-screen wrapping. A filename cannot explain a mixed-language source track to every player; choose the recipient’s supported label and retain the fuller explanation in the handoff.

Finally review the entire audience track for missing speech, meaningful sounds and necessary speaker information. W3C’s caption guidance makes clear that speech translation alone is not complete accessibility captioning. Hold any unresolved passage that changes meaning, and distinguish a reviewed local file from a destination that has actually been checked.

Keep exploring.