How to review unclear speech in subtitles

Keep the uncertainty out of the spoken words

When a subtitle contains a word you cannot confidently hear, return to the recording before correcting it. Listen with context, check relevant references and ask another suitable listener if needed. If the words remain unrecoverable, follow the destination’s convention for describing that fact. Do not turn the most plausible guess into apparently verified dialogue.

Work with media you own or are authorized to process. Keep the original recording and subtitle revision, and note the affected time interval separately. W3C describes automatic captions as a starting point that needs accuracy review: a single missing negative can reverse a sentence. Fluent-looking text deserves the same scrutiny as an obvious nonsense word.

Listen to the passage before studying the guess

Replay enough audio before and after the doubtful word to hear the whole thought. First try listening without reading the draft; then compare what you heard with the text. This simple review habit helps keep the proposed word from becoming your only candidate. Check the exact final media edit and soundtrack, especially if another take or a translated audio version exists.

If your player supports slower playback, use it as an additional listening aid and return to normal speed before deciding. Comfortable headphones may help you examine a passage, but neither a playback change nor another model-generated suggestion proves that missing information has been recovered. Keep alternatives in the review note, not as a slash-separated list in the viewer’s caption.

After a bounded listening pass, identify the next useful action. That might be asking a fluent speaker, requesting the matching original audio from the owner, or accepting that a noise masks the words. Repeating the same inconclusive replay is not a reason to approve the guess.

Use a reference to check spelling, not invent speech

An approved name list, script or visible product label can establish how a term is spelled. It does not establish that the speaker used that term in this take. W3C’s transcription guidance notes that scripts often need adjustment to match the finished audio. Compare the reference with the recording, including departures, mistakes and interruptions.

Ask a precise question: ‘Can you identify the word at 00:42 in workshop-v3?’ is more useful than ‘Is this transcript right?’ Record what the answer establishes. An owner explaining what they intended to say is different from a listener confirming what is audible. If the intended instruction matters but the recording does not communicate it, the owner may need to revise the media or supply a clearly separate clarification. Captions should not silently repair the underlying message.

Distinguish unfamiliar speech from unavailable speech

A language you do not understand is not automatically unintelligible. Route it to someone who can identify and understand it before choosing a label or translation. Netflix’s U.S. English delivery guide distinguishes translated dialogue from dialogue intentionally left untranslated and asks captioners to research the language. Those are Netflix conventions; your own destination may require a different presentation.

Also separate an unfamiliar technical word, simultaneous voices, speech covered by a machine, and deliberately obscured audio. The next step differs in each case. A terminology reference may settle spelling; another listener may resolve an accent; neither can justify revealing words intentionally removed from the final soundtrack. Avoid labels that diagnose a person or speculate about why they speak that way.

Choose an honest label using the delivery guide

There is no single universal label for every subtitle file. W3C lists [unintelligible] among common transcription practices when words cannot be understood. BBC guidance instead discourages subjective unintelligibility labels and favors explaining a verified cause, such as a sound drowning out speech. These recommendations have different scopes; square brackets alone do not make a choice correct.

Preserve the words you can establish and represent only the affected passage according to the chosen convention. Do not replace an entire understandable sentence because one short part is masked. Equally, do not describe applause, machinery or overlapping speakers unless that event is supported by the audio. If the cause is unknown, ask the delivery reviewer which neutral convention to use.

A label explains a limitation in the soundtrack; it does not restore a missing instruction. If the unavailable words determine which material to use, who a statement concerns, or whether an action is permitted, hold that part of the delivery for an owner decision. A faithful caption can still accompany a recording that is unsuitable for its intended instructional purpose.

Example: a material name under a tool noise

In this fictional workshop clip, a presenter names a material just as a powered tool starts. The draft supplies ‘felt.’ A preparation sheet lists both felt and cork, so it cannot resolve what was spoken. Keep that conflict in an external note and ask for a review of the matching recording. The sheet below is a proposed manual record, not a Polyglot feature or a tested customer example.

If the word becomes clearly verifiable, correct it and record the basis. If it remains masked, choose the agreed caption label and ask the owner whether the missing instruction requires a revised recording. Do not make ‘felt’ final merely because it was the first suggestion.

Media: workshop-v3.mp4
Interval: 00:42–00:44
Draft: material named as felt
Heard: name covered
  by tool noise
Reference: sheet lists
  felt and cork
Status: unresolved
Next: owner reviews
  matching audio
Final decision: pending
Affected exports:
  source + translation

Apply confirmed corrections in Polyglot

Polyglot supports local subtitle generation, SRT and UTF-8 WebVTT import, text review and export on Apple Silicon Macs. When correcting a source cue, check any affected translations too. Source changes can mark translations for review; an already completed translation should not be assumed to contain the correction.

Keep the uncertainty note outside the subtitle text. This workflow does not assume automatic confidence scores, unclear-speech detection or audio restoration. If you request a new line suggestion, judge it against the recording before accepting it; agreement with an earlier draft is not independent confirmation. Adding a separate missing cue may require another suitable caption editor, so do not hide a sound label inside unrelated dialogue just to fit the existing cue structure.

Close the note against the exported track

Open the exported subtitle revision with the intended video. Revisit each flagged interval at normal speed, including the words immediately before and after it. Check that the delivered text contains the approved correction or label, that its timing covers the right event, and that temporary questions or guessed alternatives have not leaked into the file.

Repeat the check for every delivered language and keep the final decision with the media and subtitle revision names. Review the rest of the track for accuracy, meaningful sounds and speaker information as well. Resolving uncertain speech is one part of caption review; it is not an accessibility certification or proof that the entire recording is ready.

Keep exploring.