Subtitle rules nobody tells you: 17 characters a second, and why yours feel rushed

Auto-generated captions are usually accurate and still uncomfortable to read. The words are right, the timings match the audio to the millisecond, and watching with them on is somehow tiring. That is not a transcription problem - it is a timing problem, and it comes down to four numbers that broadcast subtitling has used for decades.

The number that matters most: 17 characters per second

Reading speed is the whole game. A caption that holds 60 characters for two seconds is asking the viewer to read 30 characters a second while also watching the picture. They can’t, so they do one of two things: pause the video, or stop reading and let the captions become noise.

The reference point most of the industry converges on is 17 characters per second for adult programming - the figure in Netflix’s timed-text style guide, and the default our readability checker flags against. Children’s content goes slower, usually around 13. Fast dialogue in a comedy can push higher for a few cues without hurting, but a file that averages above 20 CPS will exhaust anyone who actually reads it.

Here is what that means concretely. At 17 CPS:

Caption length	Minimum time on screen
30 characters	1.8 seconds
42 characters	2.5 seconds
60 characters	3.5 seconds
84 characters (two full lines)	5.0 seconds

Machine-generated captions break this constantly, because they cut on pauses in the audio rather than on how long a line takes to read. A speaker who fires off a long sentence in one breath produces one cue with 90 characters and three seconds of airtime - technically perfectly synced, practically unreadable.

42 characters per line, two lines maximum

The second number is line length. Beyond roughly 42 characters, the eye has to travel far enough horizontally that it loses the start of the next line, and on a phone the text either shrinks or wraps somewhere ugly.

The third number follows from it: no more than two lines at a time. Three-line captions cover the picture and force the viewer to choose between reading and watching. If a sentence doesn’t fit in two lines of 42, it isn’t one caption - it’s two.

When you do split, split at a grammatical joint. A line that breaks after “the” or “of” makes the reader hold an incomplete phrase in their head for the length of a line-jump:

Bad:                        Better:
We decided to move the      We decided to move
launch to March because     the launch to March

Two lines matter for a second reason on vertical video: a three-line caption is tall enough to reach into the strip the app’s own interface covers. The caption safe zone checker draws that strip over your own frame.

The fourth number: don’t flash

A caption that appears and vanishes inside a second reads as a flicker, even when the words genuinely took that long to say. The usual floor is five sixths of a second (about 0.83 s) in broadcast practice, and a full second is safer online. Short interjections - “Yeah.” “Right.” - are exactly the ones that come out too fast from automatic tools, because the audio really is that short.

The fix is not to hold every cue longer; it’s to merge. Two one-word cues 300 ms apart should be one cue that holds for a second and a half.

Why auto-captions systematically get this wrong

Automatic captioning optimises for synchronisation, not readability. It knows when each word was spoken, so it places cues to match the waveform. The result is captions that are perfectly true to the audio and indifferent to the reader:

• long sentences spoken quickly become single dense cues,
• natural pauses become cue boundaries even mid-clause,
• fillers get their own micro-cues,
• and nothing merges, because nothing is measuring reading time.

This is why “the transcript is 99% accurate” and “the captions are unpleasant” are both true of the same file.

Checking a file in about a minute

You don’t need to eyeball 400 cues. Export your SRT or VTT, drop it into the subtitle readability checker, and it lists every cue that exceeds your CPS limit, runs past your line length, uses more lines than allowed, or sits on screen for less than your minimum. Defaults are 17 CPS, 42 characters, 2 lines - change any of them if your platform has its own house style.

Then fix in this order, because each step reduces the work of the next:

1. Merge the flickers. Cues under a second that sit next to their neighbours.
2. Split the dense ones. Any cue above your CPS limit becomes two cues with a real split point.
3. Rewrap the long lines. Break at clause boundaries, not at the 42nd character.
4. Re-run the check. A clean file usually lands at 90%+ of cues passing; chasing 100% is rarely worth it, since a handful of fast cues in genuinely fast dialogue is fine.

If your captions came out of a different tool in the wrong container, the subtitle converter moves between SRT, VTT and plain text without touching the timings.

One thing the numbers don’t cover

Standards say nothing about what you caption. Two habits do more for comfort than any timing rule: drop the fillers (“um”, “you know”) unless they carry meaning, and caption sound that matters - a door, a laugh, music starting - for anyone watching without audio. Neither shows up in a CPS report, and both are the difference between captions that are compliant and captions that are actually good.