Auto captions vs an SRT you wrote
ASR is a draft. A human or edited SRT is the names, numbers and SKUs. Short-form still burns that file in; the sidecar is not the Reel.
Auto captions are a listen. An SRT you wrote is a text file you will stand behind. Names, numbers, and SKUs are why the second object exists. After you have the words right, short-form still needs a burn-in. The sidecar is not the Reel.
ASR is the draft
Automatic speech recognition turns audio into a timed guess. add_veed_captions does that listen in 165 language codes, then burns a look. BASIC at 1×, DYNAMIC at 2×. The look is not the accuracy. A Glass render of a wrong SKU is a pretty wrong SKU.
What ASR gets right is common speech in clean audio. What it misses, first and reliably:
- Names. People, places, brands. Phonetic neighbours. On mute that neighbour is the message.
- Numbers. Prices, dates, dosages, “fifteen ninety-nine” versus 15.99 versus 1,599.
- SKUs and model strings. Alphanumeric identifiers the model has never seen as a product.
- Jargon. The word you booked the guest to say.
Overlap, music under the voice, and a Zoom webinar mix make the draft worse. Language code helps the listen (es-MX versus es-ES). It will not spell the SKU.
Treat the first output as a transcript to edit. If you burn it untouched, you have shipped the draft. Burned-in captions freeze the miss into pixels. A later “quick fix” is a new render.
Forced alignment is the sibling, not the same draft. If you already know the words, alignment places them on the clock. That is how a human SRT becomes a burn without asking ASR to invent the product name again. Auto captions run listen and alignment together, which is why a wrong word and a late word look like one failure until you read the file.
The SRT is the names, numbers, and SKUs
A SubRip file is numbered cues, comma decimals, plain text. No style. No karaoke. That poverty is the point. You can read it, diff it, send it to a reviewer who does not have a timeline.
Write or edit the SRT when:
- the claim contains a price, a dose, a legal qualifier
- talent names will be searched or tagged
- the product on screen has a code that must match the box
- you will reuse the words on YouTube long-form, a course, or a second language later
QC is a list, not a vibe. Open the run-of-show or the script. Tick every proper noun and every digit. Then watch once with the sound on for sense, once with the sound off for hold and wrap. Caption timing is whether the cue meets the mouth. Reading speed is whether a human can finish it. The SRT can be word-perfect and still too fast. Split clauses in the file. Do not hope the preset will wrap prettily.
WebVTT is the same words with a different header and a full-stop decimal. Convert; do not rename. The VTT to SRT converter exists so a player is not the first thing to reject the track.
An SRT is still not SDH until you add speaker IDs and labelled sounds. It is not a translation until you write the other language. It is not a lower third. Keep those jobs out of the cue list unless they belong there.
If you do not have time to write from scratch, generate the auto transcript, export, edit the three classes of miss, then burn the edited file. That is still “an SRT you wrote” in production terms. The authorship is the pass, not whether a human typed every article.
Then you still burn in for short-form
TikTok, Reels, and Shorts will not play your sidecar in the mute feed. The player overlay some apps offer is not your type, not your placement, and not a guarantee. Pixels are the format.
So the pipeline is:
- Lock the cut.
- Get words: ASR draft or a script.
- Make them true: names, numbers, SKUs.
- Make them readable: two lines, finishable hold, sentence case.
- Burn into the 9:16 file. BASIC for ads and courses; DYNAMIC only if the organic hook is a spoken phrase you can still read.
- Keep the SRT next to the project for YouTube long-form, the course player, and the next recut.
YouTube landscape can ingest the sidecar. Use it there. Do not paste the burned Reel into a 12-minute slot and call the captions done — and do not upload a silent TikTok with only the SRT and call the feed captioned.
Style lives in the burn, not in the SRT. Caption styles and the VEED 21+9 list are drawings. Karaoke word-highlight needs word-level times; a line-level SRT will not become Snap because you wished it. If you need pop, you need a timed kind on the burn path, after the words are right.
A recut after burn-in throws the pixels away. If you expect trims, keep the SRT as source of truth and recaption. That is cheaper than treating the MP4 as the caption file.
Pick the object per destination
| Destination | Words live in | What you ship |
|---|---|---|
| Reels / TikTok / Shorts | Burn-in | Edited transcript as pixels |
| YouTube long-form / course | Sidecar | The SRT or VTT you will stand behind |
| Paid 9:16 | Burn-in + overlay | Speech from the SRT; legal line authored |
| Archive | SRT | The file you can patch |
The AI caption generator and add captions to a video are the burn. They are allowed to start from a listen. They are not a reason to skip the read. Add subtitles automatically is the same engine with a BASIC, readability-first intent — still a draft until names and numbers pass.
Auto versus authored is not a brand preference. It is whether the SKU is right. Then the feed still wants pixels.
Edit the SRT, then burn the short in Versely.
FAQ
Can I upload an SRT instead of burning captions on TikTok?
Not if you want the mute feed to show your words as designed. Burn-in is the short-form object. Keep the SRT for long-form, courses, and recuts.
Is a cleaned auto transcript “an SRT I wrote”?
Yes, if you actually edited names, numbers, SKUs, and holds. Authorship is the pass. Shipping the raw listen is still auto captions.
Does a human SRT need a new transcription to look like Glass or Snap?
No. The words and times are the SRT. The look is the burn. DYNAMIC and word-highlight still need a finishable line; a perfect SKU in a 90-character cue will flash.