You can auto generate captions for video free in three ways: a browser tool that runs Whisper on your own machine, a desktop app like Whisper.cpp, or platform-native auto captions in editors such as CapCut or Clipchamp. The catch is where your video bytes go: most browser editors send footage to their servers, which is a non-starter for deposition clips, patient footage, or source-protected material. This walkthrough compares the free methods on accuracy, privacy, and watermark policy, then gives exact steps from loading footage to exporting an SRT file.
You have three realistic routes: platform-native auto captions on CapCut or Clipchamp, free tiers of editors like VEED.io, or a Whisper-based tool that runs transcription in your browser. All three cost nothing, but they differ sharply on privacy and output.
Platform-native captions are fast but tie you to that platform's export. Free editor tiers often add watermarks or limit caption length. Whisper-based tools give you an SRT you own, with support for roughly 90 to 100 languages and strong accuracy on clear speech, per OpenAI's Whisper documentation.
If your footage is confidential, the deciding factor is where processing happens. AetherCut runs Whisper locally in your browser, so nothing leaves your machine.
Auto captions are the machine-generated transcript itself. Subtitles are that text packaged as a separate file, usually SRT, that travels alongside your video as a sidecar. Burned-in text is rendered permanently into the frames, so every viewer sees it whether they want to or not.
The distinction changes your export choice. Sidecar SRT keeps the original footage untouched, works with platform caption players, and stays searchable and editable later. Burned-in text is the right call when the destination strips sidecar files, like some social feeds and internal LMS players, or when you need guaranteed rendering on locked-down devices.
In AetherCut you can generate captions once and export either way, or both. We cover the SRT export path in our guide to zero-gigabyte handoffs.
Two methods process everything on your machine: a desktop app running Whisper locally, and a browser tool that runs speech recognition through WebCodecs and WebGPU with the file staying in the tab. Everything else in the free category, including Kapwing, VEED.io, Clipchamp, and CapCut, moves footage to remote infrastructure to run transcription. For deposition video, patient footage, or source-protected interviews, that transfer is the compliance failure, not the captioning.
| Method | Where transcription runs | Footage leaves your machine? |
|---|---|---|
| Browser Whisper tool (client-side) | Your GPU, in the tab | No |
| Desktop Whisper app | Your CPU/GPU | No |
| Free online captioners | Vendor infrastructure | Yes |
Browser editors need shared state for collaboration and heavy transcoding, so source clips move to their servers before captions exist. Most browser-based editors process footage on remote infrastructure, including Kapwing, VEED, Clipchamp, and CapCut. If your footage falls under HIPAA, attorney-client privilege, or an NDA, that architecture alone disqualifies them.
Open DevTools, go to the Network tab, load a clip, and run transcription while watching for outbound requests carrying video bytes. Any POST with a multi-megabyte payload to a foreign domain means the footage left. Caption output itself is small text: SRT and the W3C's WebVTT format are plain-text cue lists, not media (W3C WebVTT specification). AetherCut publishes the same test: open DevTools on any tier and confirm zero outbound video bytes, then export SRT locally. It is honest about its limits too: transcription quality still depends on audio, and heavy 4K timelines run better on capable hardware, as our laptop picks for video editing cover.
On clear, single-speaker audio, Whisper-class transcription typically produces captions you can use with light cleanup: fixing a proper noun here, punctuation there. Expect most words to land correctly, and the timing to be close enough that you rarely adjust cue boundaries.
Accuracy drops in three predictable situations. Background noise or multiple people talking over each other garbles words. Heavy domain jargon, like medical terminology or statutory language, gets transcribed as plausible-sounding nonsense. Strong accents can shift word choice even when the audio is clean.
If you edit deposition footage, patient training material, or source-interview tape, manual review is non-negotiable, not a nice-to-have. A mis-transcribed name or drug dosage in captions is the same liability as a mis-transcribed transcript, and opposing counsel will read it that way. Build the review pass into your workflow: generate the captions, then read every line against the audio before export. AetherCut's lawyer-focused workflow assumes that review pass rather than skipping it. Treat any auto caption output as a first draft, never a finished record.
Open AetherCut in Chrome or Edge, then drag your file onto the timeline. The File System Access API reads the clip directly from your disk, so no bytes travel to any server. Open DevTools, check the Network tab, and you will see zero outbound video traffic.
Click Captions, then Generate. A Whisper model runs locally through WebGPU and produces a transcript synced to the timeline, usually within a couple of minutes for a short clip. Every caption appears as an editable block: fix a misheard term, split a block at the playhead, or drag its edges to retime it. Deposition and patient-footage terminology is where Whisper stumbles most, so budget a few minutes of manual review even on clean audio.
The reason this matters for regulated work is simple: correction happens on the timeline, not in a chat box. You are not pasting a transcript into a third-party service for cleanup, and you are not trusting a vendor's retention policy with privileged material. The model weights sit in your browser's cache after the first run, so repeat jobs work offline. If your footage includes names, case numbers, or diagnoses, nothing about that text ever leaves the machine, which is the whole point of client-side captioning.
That also changes how you collaborate. A paralegal can review the corrected SRT before it is attached to the case file, because the file itself is the deliverable, not a link into a vendor's dashboard. Retention policies, breach notifications, and data-processing agreements all become irrelevant when no third party ever holds the footage or the transcript. The compliance conversation shrinks to your own machine's security, which you already control.
Choose Export Captions for a plain .srt file, or enable Burn-In to render text into the frame. SRT works on YouTube, TikTok, and Instagram, per their caption format help documentation. For platform-native captions, import the SRT; for burned-in text, export your finished video. Related workflows are covered in our video editing services guide.
An SRT file is plain text: a numbered cue, a timestamp range like 00:00:03,000 --> 00:00:06,500, and the caption line beneath it. That structure is why every major platform accepts it, and why you should always export one alongside your video even if you plan to burn the text in.
In AetherCut, open the Captions panel, review the auto-generated cues on the timeline, then click Export and choose SRT. The file downloads straight from the browser tab; nothing goes to a server first.
Attaching it per platform:
Keep the SRT as your editable master. Re-export the video whenever captions change, and you never re-transcribe.
Most free routes fall into two camps. Platform-native captioning (CapCut, Clipchamp, Kapwing, VEED) generates captions at no cost but usually stamps exported video or restricts watermark-free export to paid tiers; check each vendor's current terms before relying on it, since free-tier limits change often. Desktop exports from Whisper-based tools carry no watermark because there is no vendor in the loop.
AetherCut adds no watermark on any tier, free included. Captions export as SRT or burn into the frame, and your choice of typeface, size, and position applies either way.
Where it falls short: caption generation runs locally through Whisper, so it is slower than a vendor's GPU farm on long files, and styling options are narrower than Descript's. If you need rapid turnaround on hour-long footage, a watermarked free tier may still be the faster trade.
Free captioning has real constraints, and honesty about them matters more than a feature grid.
Local Whisper transcription runs on your CPU or GPU, so speed depends on your hardware. A 20-minute deposition clip might take minutes on a modern machine with WebGPU acceleration and far longer on an older laptop. Browser-based tools also cap practical project length by memory, not policy: a 90-minute multi-clip timeline is a different job from a 5-minute trim.
Language coverage varies. Whisper handles dozens of languages well but weakens on heavy accents, code-switching, and specialized vocabulary like medical or legal terms, so expect manual cleanup in regulated work.
Formats are the quiet limit. Some free methods export SRT only, forcing you to burn in text elsewhere; others burn in but give no editable file.
AetherCut's own shortfalls: transcription quality matches the local Whisper model you run, so accuracy tops out where that model does, and very long multi-clip projects are bounded by your machine's RAM rather than a server's.
Not without a review pass. Whisper-class transcription handles clear, single-speaker audio well, but deposition and clinical material contains proper nouns, drug names, and statutory language that frequently come out as plausible-sounding errors. Generate the captions locally, then read every cue against the audio before export. A mis-transcribed dosage carries the same liability as a mis-transcribed transcript.
Yes, if you pick the right method. Desktop Whisper apps and browser tools that run transcription client-side need no sign-in, since there is no vendor account to gate the feature. Free tiers of online captioners like Kapwing or VEED require accounts, and their watermark and length limits are tied to that login. Local tools skip all of it.
Local methods do. A desktop Whisper app runs entirely on your CPU or GPU, and browser-based tools load the model once, after which the weights sit in your browser's cache and repeat jobs run offline. Platform-native captioning in CapCut or Clipchamp and free editor tiers both need connectivity, because transcription happens on remote infrastructure.
Edit directly on the timeline rather than re-running transcription. In AetherCut, every caption appears as an editable block: click the text to fix a misheard term, split a block at the playhead, or drag its edges to retime the cue. Correcting in place takes seconds per error. Pasting a transcript into a separate cleanup tool is slower and sends privileged text elsewhere.
YouTube accepts SRT sidecar files, but TikTok and Instagram do not for short-form content. TikTok's native caption tools only support text rendered into the frame, and Instagram has no SRT support for Reels, so burn captions into the video before export. Keep the SRT as your editable master regardless, and re-render whenever the text changes.
Thanks for rating!