Transcript-based editing can cut the video through its spoken words. Before deleting a passage, preserve an editable copy and check the selected range. Menu names, languages, entitlements and platform support vary by version, account and region.
CapCut Transcript-Based Editing is part of CapCut's broader AI tools workflow: it turns speech into a searchable text view and links the text changes you select to the matching video or audio. To use it safely, duplicate the project, generate the transcript in the right language, fix names and technical terms, remove only passages you've reviewed in context, and replay every cut that results.
Critical distinction: fixing transcript text can update captions, but deleting transcript words can remove the matching spoken section from the timeline. Before you confirm a bulk change, read the command and preview the timeline.
What transcript-based editing actually does
CapCut describes text-based editing as a way to change video or audio by editing a synchronized transcript. The current official pages mention keyword search, text correction, filler-word detection and removal of speech gaps. Those controls are related, but they don't have the same effect on your timeline.
- Search: jumps to a spoken word or phrase.
- Correct text: fixes the written transcript or linked caption wording.
- Delete spoken text: can cut the corresponding media range.
- Remove filler words or gaps: proposes edits that still need human review.
If you only need to create or fix subtitles, start with the auto-captions guide. Transcript editing pays off when you also want the written speech to help shape the timeline.
Prepare the project before generating a transcript
Start by protecting the original edit. Duplicate the project or timeline where that's supported, and identify the main speech track. Mute music or overlapping audio if the platform allows it, choose the spoken language, and keep a short section handy for checking names, product terms and punctuation.
CapCut’s caption guidance names accents, fast speech, background noise and low audio quality as recognition problems. Overlapping speakers and unfamiliar terms deserve a careful check too. An inaccurate transcript still works for moving around the timeline, but it's dangerous as an automatic delete list.
A review-first transcript workflow
- Import or open the speech video and place it on the timeline.
- Open Transcript-Based Editing from the current Web or Desktop route shown in your build.
- Select the recognition language and generate the transcript.
- Correct names, numbers, negations and technical terms before structural edits.
- Search for the section you need, then listen to the full sentence around it.
- Apply one small deletion and inspect the timeline result.
- Replay audio and picture together before accepting the next batch.
The official Web tool page says to right-click timeline media and choose Transcript-based editing. The Desktop guide points to Menu → Layout → Transcript-based editing and covers keyword search. Those are documented routes for different editors. They don't prove every build has the same panel, so use whichever route your version offers.
If the panel only deletes phrases or lines
Don't assume every build lets you select individual words. Before a longer edit, check whether the panel selects words, phrases or whole lines, and read the delete command. If the range you can select includes words you need to keep, cancel that deletion and make a smaller manual cut on the timeline where you can. Fixing caption spelling is a separate job, and it shouldn't require deleting the spoken passage.
Treat filler-word detection as suggestions
I'd review every proposed filler-word cut before accepting a batch. Cleaner speech only helps if the sentence keeps its meaning and rhythm. CapCut lists examples such as “um,” “uh,” “like” and “you know.” A detected word isn't automatically disposable. “Like” may introduce an example, a pause may carry emotion, and a repeated word may be emphasis, not a mistake.
| Candidate | Review | Risk after removal |
|---|---|---|
| Isolated “um” before a clear sentence | Listen through the transition | Small click or abrupt visual jump |
| Long pause | Check whether it creates emphasis | Rushed pacing |
| Self-correction | Keep the final intended meaning | Broken grammar or missing context |
| Technical term marked wrong | Correct text before any delete | Removal of essential content |
| Overlapping speakers | Use the timeline and audio, not text alone | Cutting another speaker |
Check picture continuity, not only clean audio
A transcript edit can remove words cleanly and still leave a visible jump in the speaker's face, hands or background. Watch each cut at normal speed and again frame by frame. B-roll can cover an audio edit you need, but it shouldn't hide a broken sentence or a wrong caption.
Once the structural edit is done, go back to caption segmentation and timing; the caption timing guide covers that pass. If recognition fails or captions disappear, work through the captions troubleshooting guide before you regenerate everything.
How to recover from an over-aggressive text edit
Undo right away if the timeline still holds the correct history. If you've already accepted several changes, compare against the duplicate project, restore the affected sentence and make smaller edits. Don't try to rebuild a long interview from memory after a bulk deletion.
- Keep the untouched project until delivery.
- Apply changes in small batches.
- Save a separate copy before filler-word cleanup where your version supports it.
- Review the first and last word around every cut.
- Export captions separately only after the visible text is final.
For a separate subtitle file, follow the SRT export guide. A clean transcript view doesn't by itself prove the exported subtitle file is correct.
Before delivery, replay the first and last words around each cut, check the picture for jumps, and compare the final captions with the speech. Confirm the language options and selection controls in your build before starting the next edit.