Blog

Subtitles Have a Speed Limit. Live Captions Can’t Do the Division.

November 18, 2026  ·  Tablingo

Netflix caps subtitles at 20 characters per second. Live captions can’t reach that number, because the bottom half of the fraction isn’t theirs to control.

You are watching a livestream with the captions on. The text keeps moving. You read the first half of a line, glance up for a second to see what the person on screen is doing, look back — and the line is already gone, replaced by the next one. You conclude that you are a slow reader. You are probably not. The captions were never given enough time to be read.

This is not the same complaint as delay, which has a piece of its own here. Delay is about when the text arrives. This is about how much of it arrives per second — density rather than timing — and about a number the subtitle industry settled on decades ago that live systems cannot reach.

The industry does have a number

Subtitle work looks like taste until you read a delivery spec. Netflix publishes a style guide for every language it ships, and the English (US) version puts a hard upper limit on reading speed: 20 characters per second for adult programming, 17 for children’s. The same two figures appear twice in the document — once in the section for SDH, the deaf and hard-of-hearing track, and once in the section for interlingual subtitles, the translated ones. Characters are counted with spaces and punctuation included.

The same guide caps a line at 42 characters and a subtitle at two lines. Its general requirements set a floor and a ceiling on time: an event must stay on screen for at least five-sixths of a second, and no longer than seven seconds.

Those figures are not decoration. They interlock. A full two-line subtitle is 84 characters, and at 20 characters a second, 84 characters need 4.2 seconds of screen time. That sits inside the 0.83-to-7-second window with room to spare — which is the point. The timing rules were written so that the reading-speed rule could be met. Fill both lines and you must hold them for four seconds; the seven-second ceiling stops you from simply parking the text there and calling it solved.

One caution before treating 20 as gospel: the guide does not agree with itself across languages. The English (US) page gives 20 and 17 in both of its sections. Some localised versions of the same document split the number differently — the Afrikaans page, for instance, allows up to 20 characters per second for SDH but 17 for interlingual subtitles, with no note in the English page explaining why. The number is a house convention, and house conventions drift at the edges.

For a finished file, both halves are editable

Characters per second is one division: characters on screen, divided by seconds on screen. In a prerecorded file, both halves belong to the person making the subtitle. They decide how many words go on the line, because they can rewrite the sentence. They decide how long the line stays, because they can see the silence that follows it.

This is why the style guides read like editing rules rather than timing rules. Netflix tells subtitlers to break lines after punctuation, before conjunctions and before prepositions, and never to separate a noun from its article, a first name from a last name, or a verb from its subject pronoun. Every one of those instructions exists because where a line breaks changes how fast it reads. The subtitler is not transcribing. They are laying out. And laying out requires the whole sentence in front of you before you decide where it bends.

Hence the licence the guide grants itself. Truncating dialogue is permitted — but only “where reading speed and synchronicity to the audio are an issue.” The escape hatch is defined in terms of the number it is trying to protect.

Live, the denominator is not yours

Now hand the same division to a live system.

The numerator belongs to whoever is speaking. Every recognised word goes up on the screen, and the system cannot drop three filler words to buy room, because it does not yet know which words are filler — they are still in the speaker’s mouth.

The denominator is worse. In roll-up captioning, the style most live tools use, each new word pushes the line upward until it leaves the top of the box. A line’s time on screen is therefore set by how fast the next words arrive, not by how long the line needs. When someone speeds up, both halves of the fraction move the wrong way at once: more characters come in, and each line’s dwell time shrinks. The characters-per-second figure climbs faster than the speech does.

The alternative is to wait for the sentence to finish before showing it. Then the line gets a deliberate dwell time and the division can be done properly — but waiting is delay, and delay is the failure we started by setting aside. Three things a viewer wants cannot all be had at once, and each way of captioning live gives up one of them — which one is a product decision, not a law of physics.

ApproachKeepsGives up
ScrollTimely, completeReadable
TrimTimely, readableComplete
WaitComplete, readableTimely

A human captioner with a stenotype gets closer to all three than a machine does, because a person can compress a sentence while it is being spoken — deciding on the fly which words carry the meaning. That is a skill, it is expensive, and it is why live human captioning is priced the way it is.

The rules only ask whether captions exist

You would expect a limit this well documented to be enforced somewhere. It is not.

In the United States, the captioning rule for television says captions “shall be displayed on the screen at a speed that permits them to be read by viewers” — 47 CFR § 79.1(j)(2)(ii). Read it twice. It asks for readability and supplies no number, no measurement, no test. Its reach is also narrower than the sentence sounds: the section is built around television broadcast stations, cable and satellite distributors, and other services delivering programming to homes, and it does not name internet video.

The web standard is blunter. WCAG Success Criterion 1.2.2 requires that “captions are provided for all prerecorded audio content in synchronized media.” That is the entire requirement. There is no characters-per-second figure in it, and none in the WAI’s explanation either. What the explanation does do is list among its failure conditions “captions omitting some dialogue or important sound effects.”

Put those two facts side by side. The standard treats a missing line of dialogue as a failure, and says nothing about how fast the remaining lines may move. The number that decides whether a line can be read at all lives in a streaming company’s style guide, where nobody can enforce it.

Live content has its own criterion, 1.2.4, and it is thinner still: captions “are provided for all live audio content in synchronized media.” One note in the WAI’s explanation deserves quoting in full, because it draws a boundary most people miss. The criterion, it says, “was intended to apply to broadcast of synchronized media and is not intended to require that two-way multimedia calls between two or more individuals through web apps must be captioned regardless of the needs of users. Responsibility for providing captions would fall to the content providers (the callers) or the ‘host’ caller, and not the application.”

Read the last clause again. Under WCAG, a meeting platform is not the party responsible for captioning its own meetings. A browser extension that adds captions is, in the standard’s terms, an accommodation a person arranges for themselves — not a conformance fix for the organisation they work for. Those are different things, and only one of them shows up in an audit.

A test that takes one minute

The gap is easy to see for yourself, and you need no tool to see it.

Take one minute of any captioned livestream. Count, roughly, the characters that scroll past. Divide by sixty. If the answer is above 20, the captions were asking you to read faster than a professional subtitle file is allowed to ask.

There is a simpler version that needs no arithmetic at all. When a line disappears, notice the last word you actually read. If it was not the last word on the line, the line was too fast — for you, on that screen, at that moment. That is the only measurement that means anything, and no style guide can make it on your behalf.

What live captions are for

Here is where we have to be straight about what we sell.

Tablingo reads the tab you are already playing and puts a translation under the line. It is a live tool, so it lives inside the trilemma above. It will keep up with speech. It will not hand you a text you can read at your own pace, because the thing that makes it live is the same thing that stops it from waiting.

A caption that keeps up and a text you can read are two products built on opposite choices. If you need the second one, you want a transcript, and a transcript of a livestream mostly arrives after the stream ends. If you need to follow along now, the compromise is the price of being live.

That split is not a marketing distinction. It is the same split that runs through the word caption itself: one track for people who cannot hear the audio, another for people who can hear it and cannot follow it — two products, one name. Reading speed is simply where the second product hits its ceiling.

Knowing which one you are looking at is the useful part. A caption that moves too fast is then a limit of the format, not evidence that you read badly.

If you want to try ours, the extension is called Tablingo — it reads the tab you are already in and puts the original line above the translation. Why two tools both called real-time can sit four seconds apart is here, what a caption tool actually reads is here, and the language a caption commits to before its first word is here.