Should subtitles keep filler words and repetitions?

Keep the words unless there is a reason to edit

You can sometimes omit an empty hesitation or an abandoned restart from web captions, but there is no safe rule to delete every “um,” repeated word or unfinished phrase. First check the destination’s requirements and listen to what the delivery communicates. Preserve anything that changes the meaning, certainty, emphasis or character of the speech.

This guide concerns same-language captions for recordings you own or are authorized to edit. Translation requires a separate language review. Keep the original recording and an untouched subtitle export before changing text. A shorter caption is useful only if it still gives the viewer a faithful account of the relevant sound.

Choose the editing rule for this deliverable

Write down whether the recipient wants a close record of the spoken delivery or permits limited condensation for readability. Labels such as “verbatim” and “clean” can be interpreted differently; specify how to handle fillers, repeated phrases, false starts and meaningful pauses. A student language assessment, performed scene and practical tutorial can need different treatment even when the words are identical.

W3C’s transcription guidance permits leaving out non-substantive speech in most web content while requiring accurate, honest representation. The BBC prefers verbatim wording when time and visual constraints allow it. Netflix’s English SDH guidance limits truncation to reading-speed or synchronization problems and rejects paraphrasing. These are scoped instructions, not interchangeable presets or evidence that Polyglot certifies a delivery.

If the recipient requires every spoken hesitation, retain it. If requirements are unclear and the performance itself is being judged, leave the words in until the recipient clarifies the rule. Do not quietly apply a tutorial’s cleanup convention to an audition or assessment.

Separate recognition errors from spoken repetitions

A generated draft may repeat a word that the speaker said only once, or omit a repetition that was clearly spoken. Correcting those mismatches restores the recording. Removing an actual spoken repetition is an editorial decision and needs a different justification.

Listen to the whole sentence and the response that follows. A restart may be empty searching for a word, but it may also show a change of mind. A repeated “no” can be a forceful refusal. A pause before agreement may explain another person’s reaction. Do not use a find-and-delete list of filler words: the same expression can serve different purposes in adjacent cues.

Keep grammar and dialect faithful to the voice. W3C cautions against correcting the speaker’s grammar; the BBC and Netflix also emphasize preserving speech style. Fix a transcription error without rewriting the speaker into a more formal, fluent or certain person.

Try a meaning test on a concrete example

The following invented workshop lines show three different decisions. They are text examples, not model output or a measured recording. Assume the owner permits limited removal of non-substantive fillers for an ordinary instructional video. Under a strict verbatim brief, keep the audible filler in the first line as well.

In the first pair, removing “um” may be reasonable if listening confirms it is merely a brief empty hesitation. In the second, “I think” tells the viewer that the speaker is uncertain; removing it strengthens the claim. In the third, the repeated refusal carries emphasis. Neither uncertainty nor emphasis becomes disposable because it occupies screen space.

Possible limited edit:
“Um, put the blue sheet here.”
“Put the blue sheet here.”

Keep the uncertainty:
“I think this is final.”
Not: “This is final.”

Keep the emphasis:
“No, no, leave that one alone.”

Shorten only after checking the surrounding cues

For every proposed deletion, ask whether the viewer would infer the same event, attitude and level of confidence. Check names, negatives, conditions, contrasts and corrections especially carefully. A short word such as “but” can carry the connection between two thoughts. Read the previous and following cues so an edit does not remove the setup for a reply.

If faithful text remains too fast, check timing and segmentation in an appropriate editor before cutting more. A line break changes layout, not the time available to read. Do not extend a cue into unrelated speech or remove a whole meaningful sentence to satisfy a counter. Use the separate readability guide for diagnosing those constraints.

An intentional condensation should still follow the audible sequence. Avoid displaying the conclusion before the speaker reaches it, especially when a hesitation or self-correction is important. A file can preserve its timestamps and still misrepresent the moment if its words have been rewritten too aggressively.

Keep a small decision log outside the captions

For passages that require judgment, record the video revision, cue identifier or time interval, audible wording, proposed edit, reason and review status. Keep alternatives and approval notes outside the exported subtitle text. This is a manual editorial note, not an automatic Polyglot report.

Use specific reasons such as “empty hesitation; limited condensation permitted” or “retain: repetition signals refusal.” “Sounds better” does not explain whether the viewer loses information. If the reason remains uncertain, preserve the wording and mark the decision pending. A second reviewer should be able to find the exact passage and understand the choice without guessing.

Media: paper-workshop-v3.mp4
Passage: 00:18–00:21 / cue 7
Proposal: omit initial “um”
Reason: empty hesitation;
brief permits this omission
Check: meaning and delivery
remain unchanged
Status: pending review

Review the saved export against the recording

In Polyglot, you can edit existing timed text and export a separate SRT or WebVTT file. The app preserves completed translations rather than silently replacing them when you change source text; affected translations need review. Resolve the source wording first, then check each translated version for the same certainty, emphasis and intent.

Keep the earlier export while checking the new one. Play every changed passage at normal speed with the matching video, including its neighboring cues. Reopen the exported file and confirm that the intended text changes are present, the timing remains appropriate and no unrelated cue disappeared. Then check the selected track in the receiving player.

Approve the actual file and media revision, not a general promise that the captions are “clean.” If the recipient has not accepted the editing rule or an important change cannot be justified, hold that revision. The aim is readable access to the speaker’s message, with enough of their voice left intact to understand how they said it.

Keep exploring.