A transcript is not an SRT

A transcript is words. An SRT is numbered cues with start and end times. Burn-in is those words drawn into pixels. Get a video transcript (transcribe_audio, Cartesia ink-whisper) returns editable text. It does not return a caption track. If there are no timestamps, it is not captions. Do not paste a memo into a Reel and call the job done.

This is the object split, as a procedure. It is not a sidecar-format bake-off.

Transcribe when you need words you can edit

Use the transcript job when the product is a document. For a short English clip, the free video to text converter is a free way to get that document in your browser (up to 10 minutes, no upload).

Good reasons to transcribe:

You will paste quotes into a landing page.
You will correct names and numbers before anyone styles type.
You will translate a script, then dub.
You will search an archive of spoken ads.

Procedure:

Point get a video transcript at the audio (pulled from the video, or a wav). This tool wants audio, not a hope that a video URL will come back captioned.
Pass a language code if you know it. Regenerating because you forgot language is a second bill for a field the first call accepted.
Run Cartesia ink-whisper. Read the text. It will be wrong on SKUs, numbers, and homophones. That is expected. The point of a transcript is that you can fix it in a doc.
Save the corrected document next to the video. This is not an SRT. Do not rename transcript.txt to captions.srt.
Stop if a document was the whole brief. Do not burn type “while you are here” unless the video will ship with words on screen.

Accuracy follows the mix. Overlap, beds, and heavy room noise degrade the page. Isolate first if you needed a clean document from a smash. Isolation is a different task.

What this pass will not give you:

In and out points per line.
Speaker IDs unless you type them.
A file a player can toggle.
Styled, word-timed type on the picture.

The neighbouring agent job is the same split: transcribe-audio-to-text is words. Captioning is a different call.

Build an SRT only when you need cues

SubRip is the interchange format: index, timestamp pair with a comma decimal (00:00:01,000), text, blank line. No official colour, no position, no per-word highlight. Karaoke in an SRT is a misunderstanding. If a client asked for “styled SRT,” they usually want a burned MP4 plus a plain sidecar for archive, or they mean WebVTT.

Procedure when you actually need a sidecar:

Start from the corrected transcript, not from a second raw ASR pass you will disagree with later.
Split into cues at clauses, not in the middle of a name. Latin-script lines commonly sit near 32–42 characters and two rows. That is a guideline, not a law the file format enforces.
Attach start and end times that match the mouth. A transcript with times you guessed in a spreadsheet is a fanfic SRT. Get times from a captioner, a forced-aligner, or the burn tool’s own pass — then export.
Keep the comma decimal if the destination is SRT. WebVTT uses a full stop and a WEBVTT header. Renaming .srt to .vtt is how players reject the track.
Treat the sidecar as words and times, not as the look. Short-form feeds will not draw it.

Checkable SRT test:

Open the file in a text editor. Do you see --> and timestamps. If you see paragraphs and no clocks, it is still a transcript.
Load it against the video in a player that accepts sidecars. Do lines arrive with the mouth. If you cannot load it because there are no times, you do not have an SRT.

Versely’s in-browser converters exist for VTT/SRT when a client asked for a track. The slug on that free tool is vtt-to-srt-converter. A burned Reel is still a different deliverable.

A constant offset fixes a track that is uniformly early or late (you trimmed an intro). Progressive drift is a frame-rate mismatch. Do not “fix” drift with an offset.

Burn-in when the feed will never enable a track

Mute autoplay will not open your sidecar. Burned-in captions are pixels. That is the short-form object.

Procedure:

Picture lock. A later trim is a recaption.
Audio lock. Caption what is spoken now.
Run add captions to video. It transcribes the video’s soundtrack and burns styled, timed lines. 165 language codes. BASIC 21 looks at 1×, DYNAMIC 9 looks at 2×.
Correct names and numbers on that pass the same way you would on a transcript. The generator will not know your SKU.
Preview the look on the first seconds. Then burn.
Watch muted at phone size, in the safe area. If the line is not on screen without a CC toggle, you shipped a transcript in a folder, not a captioned ad.

Do not upload an SRT to TikTok or Reels and assume the feed will show it. The feed plays the picture.

Do not ask a video model to paint “yellow captions” into the generate. That is texture. You cannot restyle it. You cannot fix a late cue.

If the client needs both a Reel and a course player: burn the short-form file, and export a sidecar from the same corrected words for the player. Two objects, one script. The AI caption generator is the burn door.

Pick the object, then stop mixing them

A single UGC week often needs all three. Mixing them is the defect.

Need	Object	Tool
Quotes, blog, script to translate	Transcript (no times)	`get-a-video-transcript` / `transcribe_audio`
YouTube long-form toggle, freelancer handoff, archive	SRT or VTT (cues)	Caption pass + sidecar export / converter
Reels, TikTok, Shorts, paid mute-autoplay	Burn-in (pixels)	`add-captions-to-video`

Procedure for a file that will do all three:

Transcribe. Correct the document. This is the source of truth for words.
Burn from the locked video (the tool will ASR again unless you are in a path that accepts a known transcript — still spot-check that the burn matches the doc).
Export a sidecar if someone asked for cues. Do not deliver the .txt and call it an SRT.
Do not paste the transcript into Instagram alt text and call that captions. Alt text is a different object.
Do not paste the transcript into YouTube’s description as fake chapters. Chapters are a different object too.

If you only needed a memo, you already stopped at step 1. Paying for DYNAMIC burn-in to get a paragraph you will email is the wrong finishing spend.

Words first, cues only when a player needs them, pixels when the feed is mute. Run the split in Versely. Bring a video URL and a name for the object you actually need.

FAQ

Does get-a-video-transcript add subtitles to the file?

No. Plain text out. Burn-in is add-captions-to-video. An SRT is a sidecar you only have after someone attached clocks.

Can I turn a transcript into captions by adding line breaks?

No. Line breaks are not timestamps. Without in/out points you still have a memo.

Is a VTT a transcript?

No. WebVTT is a cue file with times (and optional position). A transcript is the words with the clocks stripped or never added.