“15% off” appears as “50% off” in the auto-captions, and the video is already queued for publication. One misrecognized digit turns a subtitle into a false price promise for viewers watching without sound. That is more consequential than a missing comma.
Using a fictional 90-second event video, this guide shows how to generate and review auto-captions in CapCut (the international product) and export an SRT file. It is aimed at short-video teams, course editors, and anyone delivering captions alongside a web video. AI creates a draft; people must verify facts and timing against the original audio.
List the risky facts before editing
Suppose the presenter mentions an event called “Autumn Experience Week,” an October 18 registration deadline, a 15% discount that cannot be combined with other offers, and a contact named Zhang Ning. Numbers, negation, and names deserve special attention because errors can change what viewers believe. Make a small fact card with approved spellings and terms, and share it with the editor. A discount that has not been approved should not become a definitive claim in the captions.
Keep the source video untouched and work from a copy. Record the video version, language, caption editor, and intended publication date. If background music sits under narration, lower or temporarily mute it in the working copy so the recognition pass hears clean speech; do not delete the original audio. CapCut's troubleshooting guide lists noisy audio, incorrect language settings, connectivity, and muted tracks as common causes of recognition problems.
Generate editable captions, not a final render
Import the working copy into a CapCut desktop or web project. Find Auto Captions under the caption or text controls, select the main spoken language and the correct audio track, and generate a first pass. Labels vary across versions. CapCut's official guide confirms recognition on web, desktop, and mobile, while speaker differentiation and SRT export can depend on platform, region, and version. The workflow here focuses on desktop review and export rather than pretending every mobile interface is identical.
Before editing individual words, check the entire timeline. Do captions cover all the speaking sections? Was intro music or end silence turned into invented dialogue? If the whole transcript is in the wrong language or largely garbled, check the recognition source and regenerate. If only a few words are wrong, keep the existing timeline and fix those blocks. CapCut's editing help describes manual correction and splitting caption blocks.
First pass: verify facts by listening, not guessing
Open the caption editor, wear headphones, and play through the timeline. Ignore fonts and animation for now. Compare high-risk words with the approved fact card. In this fictional sample, imagine the recognizer turns “fifteen percent” into “fifty percent,” October 18 into October 8, and Zhang Ning into a different name. Replay the relevant seconds, then confirm them against the audio and approved event details. If the audio itself is unclear, mark it for confirmation rather than letting AI infer what the presenter must have said from a poster.
Pay close attention to “can be refunded” versus “cannot be refunded,” and “this week only” versus “starting this week.” A missing negation reverses the rule even when the rest of the caption is correct. URLs, model numbers, and price units also merit their own check. Keep an issue log with timecode, machine text, audio or approved text, and outcome: 00:12 / 50% / 15% / corrected. A reviewer can then audit the riskiest moments without replaying the entire video.
ChatGPT or Claude can scan the text for risky claims. Provide only caption drafts and fact cards that you are authorized to use; do not upload private customer footage without permission. Here is a prompt:
You are reviewing video captions. I will give you timecoded auto-caption text and approved event facts. List only blocks involving numbers, dates, names, products, URLs, negation, or policy promises. Output timecode | caption | concern | evidence needed | action. Do not silently replace the draft with a guessed answer. If you lack audio or approved evidence, write “human listening required.” Do not rewrite the marketing tone or remove silent sections.
AI's list is not evidence of what was spoken. Never paste a newly “polished” discount or deadline back into the captions unless the audio or approved materials support it.
Second pass: check sentence breaks and sync
After facts are corrected, focus on viewing. Automatic recognition may put 12 seconds of speech into one unreadable caption block. Split it at natural pauses and adjust start and end times so words appear with the speech they represent. Avoid showing the second half of a sentence before it is spoken merely to make the layout look neat; that can reveal an answer early and break synchronization for deaf viewers.
Watch the entire video at least once. Inspect fast speech, transitions, and music-heavy moments. Check for overlapping captions, clipped final words, and conflicts with visible buttons or prices. When meaningful non-speech audio carries information, such as an alarm or a consequential off-camera reply, caption it appropriately rather than transcribing dialogue alone. The W3C's caption guidance explains that unverified automatic captions are not sufficient for accessibility and that meaningful sound and synchronization matter.
A bilingual video needs another pass. Translate from the verified facts, not the uncorrected machine text. Keep date formats, currency, and proper names consistent. English sentence boundaries may differ from the Chinese ones; do not blindly copy three Chinese time slots onto three English lines. Complete and preview each language track separately.
Export SRT, then inspect it outside CapCut
For a website, video platform, or partner that accepts switchable captions, deliver a caption file rather than only burning words into the picture. CapCut's desktop export troubleshooting page says to select Caption Export and choose SRT or TXT. The option's availability and position can differ by platform, plan, and version. If you do not see SRT, consult the current product help instead of assuming mobile can always export it.
A minimal SRT extract looks like this. Each cue has an index, start and end time, text, and a blank line before the next cue:
1
00:00:12,000 --> 00:00:15,000
The event discount is 15% this week only.
2
00:00:15,200 --> 00:00:18,600
Registration closes on October 18.Open the exported file in a plain-text editor. Check that names and percent signs render correctly, timecodes ascend, and no cue has negative duration or a large accidental gap. If the destination platform permits it, re-import the SRT into a test project and play 20 seconds. This catches files that looked fine in the editor but shifted during handoff. Include the video version and language in the filename, such as autumn-week_v03_en.srt, rather than calling every revision subtitle-final.srt. Archive the SRT with the final video, source audio, and fact card.
Three common failures, in the right order
- No captions appear: verify the audio track, mute state, language, and network. Test a clear 15-second excerpt before processing the whole video again.
- Only a few words are wrong: correct the affected blocks and listen to surrounding lines. Do not regenerate the entire timeline and forget a second review.
- The video exports but SRT does not: confirm separate caption export, file format, and save path. Check whether the text is an auto-caption track rather than ordinary manual text. Preserve the project and a text backup before troubleshooting further.
The handoff checklist is short: source video preserved; numbers, names, and negation replayed; captions synchronized with speech; SRT opens without garbled characters; a random sample has been played back; reviewer and version are recorded. Auto-captioning saves draft typing time, not the responsibility to verify the result.