Generate subtitles from the selected final audio, correct the transcript, then time each cue against the scene’s meaning. For drama, accurate words are only the first pass: a caption can identify the wrong speaker or reveal a secret before the performance does.
Keep an editable subtitle master and a clean picture version. Burn captions into a delivery copy only after the edit is stable, unless the destination explicitly needs a separate supported caption track.
Start with one authoritative dialogue version
Create a handoff containing the video edit, selected audio, script revision, and character-name glossary. A glossary should include unusual names, pronunciation-sensitive terms, and words the recognizer repeatedly confuses.
If the final voice says “I never sent it” while the draft says “I didn’t send it,” the subtitle follows the performance. If the change alters the story, return it for an editorial decision; do not silently force the transcript to match whichever version is convenient.
This matters more in current AI workflows because dialogue can arrive from several places: a separate speech tool, native audio-video generation, a localized voice track, or a later pickup. A file labelled “final” is not enough evidence that those versions agree.
For separate tracks, transcribe the clean dialogue when available, then validate against the final mix. For a mixed native-audio file, pay extra attention to music and overlapping speakers.
Use automatic captions for the draft, not the final decision
CapCut’s official help documents Auto Captions or Recognise Subtitles and warns that options vary by platform, region, and version. Follow the controls present in your installation.
Use this sequence:
- Import the correct edit and choose the spoken language. Do not ask the recognizer to infer a language switch from a character name.
- Generate the draft. Keep a copy before extensive manual correction so recurring recognition problems remain visible.
- Correct meaning-changing words first. Negations, numbers, names, ownership, and relationship terms can change the plot.
- Assign speaker context where necessary. An off-screen line may need identification even when its transcription is perfect.
- Retain an editable export where supported. Verify the format through a round-trip import rather than assuming export success means compatibility.
The reviewed ElserStudio documentation does not establish automatic transcription and SRT export across every version. Finish captions in an appropriate external editor when your installed workflow does not supply them.
Prevent a subtitle from delivering the twist early
Consider this fictional line after a pause:
The signature … is yours.
Showing the whole sentence as soon as the speaker begins gives fast readers information before the actor reaches it. A better cue boundary may preserve the pause:
1
00:00:04,000 --> 00:00:05,700
The signature...
2
00:00:06,200 --> 00:00:07,500...is yours.
These timestamps illustrate syntax and a dramatic boundary; they are not timings measured from an existing clip. Adapt the cues to the selected performance.
Do not split every sentence this way. Excessive cue changes can make reading harder and turn ordinary dialogue into artificial suspense. Use the split where the withheld information genuinely matters.
Also inspect translated captions. Different word order can expose the reveal earlier than the source language. A literal translation with identical cue boundaries may be grammatically awkward or dramatically wrong.
Use timing diagnostics rather than a rigid character rule
For any cue, calculate:
reading load = words displayed ÷ seconds on screen
This is a diagnostic, not a universal readability standard. A dense name, unfamiliar term, or visually busy shot can require more attention than common words of the same length.
If a caption is difficult to read, choose the correct repair:
Problem
Useful repair
Repair that can make it worse
Too much text for the available time
Split at a natural boundary or revise unnecessary speech upstream
Shrink the font until everything fits
Cue appears before the relevant information
Retime the reveal
Delay the whole sentence until speech is over
Speaker becomes unclear after a cut
Add appropriate identification or adjust coverage
Assume every viewer recognizes the synthetic voice
Cue straddles an unrelated scene change
Review the exit point and audio continuation
Apply the same fixed duration to every cue
Watch at normal speed after making changes. A technically precise timestamp can still produce an uncomfortable reading rhythm.
Understand what an SRT file preserves
SRT stores cue numbers, time intervals, and text. Do not rely on it to reproduce every font, animation, or branded treatment from a caption editor.
Save in a compatible encoding such as UTF-8, then reimport the file and inspect names, punctuation, and non-English characters. Confirm that line breaks and timing survive in the actual delivery application.
If you need exact visual styling, burn it into a delivery copy while preserving the clean master and editable text. If you need viewers to switch languages or turn captions off, verify the destination’s separate-caption support. A file can be valid SRT and still be unsupported by a particular upload surface.
DaVinci Resolve’s edit documentation describes subtitle support including imported SRT. That is distinct from which edition offers automatic speech transcription. Official edit documentation.
Position captions around the actual story evidence
Keep captions away from the object viewers must inspect. If a scene turns on a receipt at the bottom of the frame, a large fixed lower-third caption can hide the evidence.
Establish a readable default position, then make deliberate exceptions. Test them with the destination interface visible; buttons and post text may cover different areas from your editor’s empty canvas.
Do not put all information into a universal “safe zone” rectangle and assume the work is finished. The usable space changes with content and interface. A practical review checks faces, hands, essential objects, and captions together.
Keep graphics restrained during consequential dialogue. Animated word-by-word highlighting can compete with a subtle reaction; it should serve the format rather than automatically accompany every sentence.
Update captions when the edit changes
Picture changes can move cues even when no words change. Replace a shot with a longer take, shift an interruption, or add a pause, and the old subtitle file may become inaccurate.
Identify the first changed edit point and inspect everything that follows until synchronization is re-established. Do not assume only the visibly replaced shot needs review.
Export a short test, play it on a phone, and check for duplicate native captions over burned-in text. Keep the cue file, language, and video version together so a later localization starts from the correct source.
FAQ
Should subtitles include stutters and false starts?
Choose a consistent editorial and accessibility approach appropriate to the delivery. Preserve meaning and important performance details; avoid automatically cleaning dialogue until its emotional hesitation disappears.
Can automatic translation handle a fictional name glossary?
It can assist, but inspect the output. Provide the approved names and terms, and have a competent speaker review wording and reveal order before release.
Should subtitles explain a silent character's thoughts?
Only when that is an intentional written or audible storytelling layer. Do not invent dialogue in the captions to explain a reaction that the scene failed to communicate.
Use ElserStudio to finish the scene and retain the selected dialogue, then complete the caption pass in the editor best suited to your delivery.