Two quite different things are called subtitles, and the difference is not how they look. One is a small text file listing what is said and when, which the player reads and draws on top. The other is text painted into the picture itself, indistinguishable from anything else in the frame.
Which to use
The separate file wins on nearly every axis. It can be turned off by somebody who does not want it, restyled by somebody who needs bigger text, corrected without touching the video, translated into six languages as six small files, and — the part people forget — read by search engines. A transcript is text, and text is the only part of a video anybody can index.
Its single weakness is decisive when it applies: a player that does not support subtitle files simply ignores it, and a good deal of social video is played in exactly those conditions. If you cannot control where the video will be played, the words have to be in the picture.
Which gives a rule that covers most cases. Your own site, YouTube, Vimeo, anything with a real player: use a file. Feeds, stories, embedded autoplay clips, anywhere silent-by-default: burn them in. And if a clip is going to both, keep the file and export a burned-in copy from it, rather than the other way round — you can always burn a file in later, and you can never get a file back out of pixels.
Why the names come out wrong
Automatic captioning is good at sentences and bad at nouns, and the reason is worth knowing because it tells you what to check.
Speech recognition is not matching sounds to words one at a time. It is choosing the most likely sequence, weighing what it heard against what tends to follow what. That is why a mumbled ordinary word usually comes out right — the surrounding sentence pins it down — and it is also why a name does not. Your company, your product, a colleague's surname, an industry term: none of them are common sequences, so the recogniser substitutes something that is.
The practical consequence: proofread the proper nouns and skim the prose. Fixing three product names does more for a viewer than fixing thirty commas, and a name repeated forty times in a talk is forty errors from one correction.
What actually degrades the transcript
- Two people talking at once. The hardest case there is, and common in interviews. Nothing fixes it after the fact.
- Music or noise under the speech. Background music that a listener tunes out is, to a recogniser, competing signal.
- Distance from the microphone. Room reverberation smears the consonants that distinguish similar words. A cheap microphone close beats an expensive one across a room.
- Accents outside the training data. Models are markedly better on the accents they saw most of, which is not a statement about the speaker.
- Very compressed audio. If the file has been through several platforms, the high frequencies that carry consonants may already be gone.
All of which point the same way as always: the recording is where accuracy is decided, and no amount of processing afterwards recovers what was not captured.
Timing matters more than people expect
A caption that appears half a second late is more tiring to watch than one with a typo in it, because the reader is continuously re-syncing. Two habits help: keep a caption on screen long enough to be read at a normal pace — a line or two, not a paragraph — and break lines where the sentence breaks rather than where the box ends.
If you trim the video after generating subtitles, the file no longer lines up: every timestamp after the cut is out by the length of what you removed. Caption last, or re-caption after cutting.
Captions are not only an accessibility feature
They are that first, and it is reason enough. But a large share of video on a phone is watched with the sound off — in bed, on a train, in an office — and a clip with no captions is a clip those viewers scroll past. The words are also the only machine-readable part of a video, which is why a platform can find a moment inside an hour-long talk.
So captions serve three audiences at once: people who cannot hear the audio, people who are not listening to it, and software that cannot listen at all.
Doing it without uploading the video
Most automatic captioning services want the video on their server, which for an unreleased cut, an internal briefing or a client's footage is a copy you did not intend to make.
The tool here runs the speech model in your browser. The first use downloads the model, which is not small; after that the transcript is produced on your own machine and the video never moves. You get the .srt to edit, which is the artefact you want to keep — and the corrections you make to it are the ones nobody else can make for you.