Why subtitles stopped being an afterthought
Something shifted around 2018. Facebook published a number that made marketing teams sit up: roughly 85% of people scroll through video with the sound off. TikTok then turned captions into a design language rather than an accessibility feature. Burned-in text became part of the aesthetic, and a video without words on screen started to look unfinished.
For anyone producing more than a couple of clips a month, the old workflow collapsed. Transcribing by hand runs at roughly five to ten times real time, so a 15-minute interview could eat two hours before you’d typed a single caption. That arithmetic is why an AI video subtitles generator went from nice-to-have to standard kit in about two years.
What the tool is actually doing under the hood
“AI subtitles” covers four distinct jobs, and tools differ wildly in how well they handle each one.
1. Transcription
Most modern tools run a speech-recognition model (OpenAI’s Whisper is the most common base, alongside proprietary engines) that converts audio to text with word-level timestamps. On clean studio audio, accuracy routinely lands in the 95–98% range. On a windy street interview with a bus going past, expect that to drop into the seventies.
2. Timing and segmentation
Raw word timestamps are not captions. The tool has to decide where lines break, how long each stays on screen, and how many words fit comfortably. Good engines respect reading speed, around 15–20 characters per second, and never leave a single orphan word dangling on its own cue.
3. Formatting
Punctuation, capitalisation, speaker labels, and the removal of filler (“so, um, yeah”) all happen here. This is where cheaper tools show their limits: they’ll hand back a wall of lowercase text with no full stops.
4. Output
Export as SRT or VTT sidecar files, or burn the text directly into the video pixels. More on that choice below.
Where AI captions hold up, and where they fall over
Give an engine a podcast recorded on a decent USB mic and it will embarrass a human typist on speed and price. The failure modes are predictable once you’ve seen a few:
- Proper nouns and brand names. “Plaud” becomes “plowed.” Product SKUs come out as random strings of letters.
- Numbers and units. “Fifteen hundred dollars” versus “$1,500” versus “1500 dollars” is a coin flip.
- Homophones. Their/there, compliment/complement, cache/cash. The model picks whichever it finds more likely.
- Crosstalk. Two people talking over each other, and the transcript turns into a merge of both speakers.
- Accents and code-switching. Heavy Scottish, Nigerian, or Indian English, or a Spanish sentence dropped mid-paragraph, will drag accuracy down noticeably.
- Music and silence. Long instrumental intros sometimes get hallucinated into words nobody said.
The ten-minute human pass
No tool I’ve tested removes the need to read every line once. Ten minutes on a 20-minute video, with a checklist, catches almost everything: names, numbers, anything inside quote marks, the first two seconds of every speaker change, and any cue sitting on screen for less than a second. Reading out loud is the fastest bug detector, because your mouth trips over the same errors your eyes skim past.
Burned-in captions or a subtitle file?
These are two different products wearing the same name. Open captions are baked into the frames. They survive re-uploads, screen recordings, and every platform’s autoplay behaviour, which makes them the right call for social. Closed captions live in an .srt or .vtt file that a player can toggle, and they’re mandatory if you publish to YouTube with multiple language tracks or deliver to a client with accessibility requirements.
Plenty of teams do both. Burn a styled version for the vertical cut, attach the sidecar file to the long-form upload.
Translation is the harder problem
Auto-translating subtitles is genuinely useful and genuinely risky. A literal translation preserves meaning while destroying timing: German runs long, Japanese runs short, and a cue that read comfortably in English can overflow the frame once translated. Better tools offer a translate-then-re-time step, and you should always have someone who speaks the language spot-check before publishing.
Dubbing raises the stakes further, because audio and mouth shapes now have to agree. If you’re heading that way, it pays to understand what separates a convincing lip sync from a creepy one. It comes down to phoneme matching and micro-expressions far more than most people expect.
The tools, briefly
Most people will meet subtitles inside a general editor rather than a standalone caption service. Kapwing handles auto-captions, styling presets, and translation in a single browser tab, and it’s forgiving if you’ve never touched a timeline before, as this Kapwing AI review lays out. VEED sits in similar territory with a stronger translation story and a stricter free tier; the VEED AI review covers where it stumbles. Standalone options like Whisper-based desktop apps or Descript appeal if you’d rather own the transcript file than rent it.
What to check before you pay
- Language coverage, and whether translation is included or a paid add-on.
- Export formats: SRT, VTT, TXT, and whether styled captions burn in at full resolution.
- Watermarks on the free plan. This is the single most common surprise.
- Per-minute limits versus a flat monthly cap. A 40-minute webinar eats a small allowance fast.
- Who owns the transcript, and whether your audio is used for model training.
Free tiers vary more than the marketing pages suggest. A practical test of how much of VEED AI is actually free found the caption feature usable but the export restrictions decisive for anyone publishing daily.
A workflow that holds up on a busy week
Upload the audio first, before any editing, and let the transcript generate while you cut. When the rough cut locks, re-run captions on the final timeline so the timing matches what viewers actually see. Style once, save it as a preset, and reuse it. Consistent caption placement is a small thing that makes a channel look deliberate.
Then do the read pass. Fix names, fix numbers, split any cue running past about 42 characters per line, and check that captions don’t sit under a platform’s UI furniture. TikTok’s right-hand buttons have eaten more captions than any encoding bug.
The honest framing is that AI solved the typing, not the judgement. It removed maybe 80% of the labour and left you with the 20% that decides whether the finished thing looks professional. Treat the transcript as a first draft from a fast, literal-minded assistant. Folding captions into a wider AI-assisted pipeline saves more time still, though it’s worth knowing what an AI video editor genuinely saves and what it quietly costs before you rebuild your whole process around one.

