Caption speed & display time: why both checks matter
A comparison of 134 selected NASA caption files shows what a duration check and a reading-speed check each find.
Influence IQ analysis of 64,347 cues in 134 English-dominant caption files for NASA videos dated 2025–2026, retrieved November 16, 2026 UTC. The thresholds are examples from Netflix’s U.S. English guidance.
Abstract
Among 125 selected files with at least 1 cue above 20 characters per second, the median file had 94.6% of its fast cues outside our short-duration flag. Among 81 files with at least 1 short cue, the median file had 50.0% of its short cues at or below the speed threshold. We used Netflix’s U.S. English adult speed and general duration limits as reference checks. These NASA files are not a sample of social videos or a test of viewer comprehension. [1][2][3][5]
You have captions on a video and a few minutes to review them before publishing. A short flash is easy to spot. A longer caption can still ask someone to read a lot of text quickly. Checking both how long a caption stays on screen and how much text it contains gives you two useful queues for the review.
Time on screen & reading speed

Two ways to flag a cue
A cue is a piece of caption text with a start time and an end time. Display time measures how long that cue is shown. Reading speed divides the number of characters by the time on screen. A duration check needs the timestamps; a speed check needs the text as well. Spaces and punctuation count in Netflix’s calculation. [9]
Imagine a cue with 24 characters shown for 1 second. Its rate is 24 characters per second, above the 20-character reference, yet its duration clears the 5/6-second reference. An 8-character cue shown for half a second has a rate of 16 characters per second, but it is shorter than that duration reference. These are worked examples, not captions from the sample. [2][3]
We flagged cues above 20 characters per second or shorter than 5/6 second, allowing half a millisecond for timestamp rounding at the duration boundary. The first threshold comes from Netflix’s U.S. English adult guidance; the second comes from its general timed-text requirements. Other delivery specifications can use different limits. [2][3]
To calculate a cue’s rate, count its characters, including spaces and punctuation, and divide by its display time in seconds. Line breaks are not characters in this calculation. A cue with a zero or negative duration needs a timing correction before its rate can be assessed. A whole-video average can be useful, but it does not identify an individual crowded cue.
What each check found
In 125 of the 134 files, at least 1 cue exceeded the speed reference. For each of those files, we calculated the share of fast cues that the short-duration check did not flag. The median of those 125 shares was 94.6%. A median describes the middle file after ordering the file-level shares; it does not mean 94.6% of all fast cues were missed. [5]
A short cue could also pass the speed check. Of the 81 files with at least 1 short cue, the median file had 50.0% of its short cues at or below 20 characters per second. The 2 medians have different bases: one uses files with fast cues, the other uses files with short cues. A file can belong to both groups. [5]
The finding was not limited to a single cue. In 124 of the 125 files with fast cues, at least 1 fast cue was outside the short-duration flag. In 68 of the 81 files with short cues, at least 1 short cue was at or below the speed reference. The files were selected from NASA’s collection, so these counts do not describe a typical creator’s video.
Each check found cues the other did not flag
For each eligible file, we calculated the share of cues flagged by one reference check but not the other. The table shows the median of those file-level shares.
| Reference check | Qualifying files | Median share |
|---|---|---|
| Fast cues outside the short-duration flag | 125 | 94.6% |
| Short cues at or below the speed reference | 81 | 50.0% |
Why the counting unit matters
Pooling the cues gives a different picture. Of 19,138 fast cues across the 134 files, 10,529 were outside the short-duration flag: 55.0%. Of 10,620 short cues, 2,011 were at or below the speed reference: 18.9%. Here each cue has equal weight, so a long recording can contribute far more than a short one. [5]
One long recording contributed 5,168 of the 8,609 cues flagged by both checks, or 60.0%. The file-level medians give each qualifying video one share in the calculation; the pooled counts give each cue one place. Both calculations describe the same selected files, but they answer different questions.
The four cue categories also make the overlap visible: a cue can be flagged by speed only, duration only, both checks, or neither. The last category means only that it passed these 2 numerical checks. It does not establish that the words, synchronization, line breaks or placement are correct.
The categories do not choose the edit for you. A cue flagged by both checks might improve if there is room to extend it, but that change still needs a check against its neighbors and the audio. A cue flagged by only one check needs review against that measure even when it passes the other.
The same cues, counted across all files
Each cue belongs to 1 of 4 groups according to whether it exceeded 20 characters per second and whether it was flagged by the 5/6-second duration check.
| Cue category | Cues | Share |
|---|---|---|
| Speed flag only | 10,529 | 16.4% |
| Both flags | 8,609 | 13.4% |
| Duration flag only | 2,011 | 3.1% |
| Neither flag | 43,198 | 67.1% |
A practical review
Start by checking that the captions contain the spoken words and meaningful sounds. Review names, technical terms, speaker changes and punctuation against the audio. Automatic captions can be a useful starting point, but they still need review. U.S. government caption guidance also asks editors to keep captions synchronized and on screen long enough to read. [6]
Next, inspect the speed and duration flags separately. A fast, longer cue may need more time or a different break. A very short cue may need to stay visible longer or be combined with a neighboring cue. Any change has to preserve the meaning and fit the audio and surrounding visuals; a numerical flag alone cannot choose the edit.
Finally, play the video with captions enabled and check them in the intended format. Look for timing, line breaks, placement and words hidden behind interface controls. In a qualitative study of TikTok captioning, participants reported captions that felt too fast; the researchers did not measure caption speed. The NASA counts cannot predict how those viewers would experience another video. [7]
After retiming, run the checks again. Extending a cue can overlap the next one or move it away from the speech. Combining cues can remove a short flash while creating a crowded line. Watch the result in the intended player and keep the words and timing together.
Review captions in context

Check the words
Compare speech, names, speakers and meaningful sounds with the captions.
Check both measures
Review reading-speed and display-time flags against the applicable specification.
Play it back
Check synchronization, line breaks and placement in the intended video.
Choose a reference for your work
The numbers in this study are reference checks, not universal pass marks. Netflix’s 20-character limit is for U.S. English adult programs; its children’s limit differs, as do its guides for other languages. A client, distributor or accessibility policy may set a different requirement. Start with the specification that applies to the work you are delivering. [2]
Changing the thresholds also changes the flagged sets. At a stricter 17 characters per second, the median of 128 files with fast cues had 96.0% of them outside the same short-duration flag. With a 1-second duration check and the 20-character-per-second speed reference, the median among 125 files was 89.3%. These are sensitivity checks on the same NASA files, not separate studies. [5]
A rate or duration limit cannot establish whether a caption is accurate, readable to every viewer or appropriate for a particular screen. Consider the words, the available time, the speaker and the visual context together. The numerical checks help locate cues for a human review.
How we analyzed the files
Mark Studios sampled NASA Image and Video Library records dated 2025–2026, selected files with SRT captions, removed duplicates and files with fewer than 5 captions containing text, and published a list of 139 retained NASA files. We retrieved those 139 files on November 16, 2026 UTC. We excluded 5 Spanish-dominant files after a language check, leaving 134 English-dominant files and 64,347 nonempty cues with positive duration. One file contains English sound descriptions rather than dialogue. [1][4][5]
We counted Unicode characters after removing formatting, line breaks and outer line whitespace. We calculated each cue’s duration in milliseconds and compared the two flags within the same cue. For the 5/6-second boundary, we allowed 0.5 milliseconds for timestamp rounding. An equality at 20 characters per second did not count as above the speed limit. [2][3][5]
We excluded 6 empty cues and recovered 1 cue whose text followed a blank line after its timestamp; a separate FFmpeg conversion treated that text as part of the timed cue. An independent parsing and counting path agreed with the final cue measurements. The source’s earlier rounded percentages can differ because its calculation treats leading whitespace and floating-point values at the speed boundary differently. Files can change after retrieval; the package records their hashes. [5]
The NASA records include selected educational and coverage videos, with caption spans from short clips to hours. They were not sampled as social posts, and the files do not show what viewers read or how the captions were made. We did not analyze the separate MIT files in the source study. Mark Studios’ measurements are released under CC BY 4.0; NASA’s media terms also distinguish its own material from third-party content. Our download contains aggregate results, source links, hashes and code, not the caption files. [1][8]
The downloads contain aggregate results, source links, hashes and code. The original caption files remain with NASA.
Sources & notes
- Mark Studios, Caption Reading Speed Study, collected October 15, 2026; per-video measurements CC BY 4.0.
- Netflix, English (USA) Timed Text Style Guide: U.S. English adult and children’s reading-speed limits.
- Netflix, Timed Text Style Guide: General Requirements, including the minimum subtitle duration.
- NASA Image and Video Library, source of the selected caption files.
- Influence IQ aggregate results, source-file hashes and reproduction code.
- U.S. General Services Administration, Captions and Transcripts, updated August 2025.
- McDonnell and colleagues, “Caption It in an Accessible Way That Is Also Enjoyable”: Characterizing User-Driven Captioning Practices on TikTok, CHI 2024.
- NASA Images and Media Usage Guidelines, including conditions and third-party rights.
- Netflix, How is reading speed measured? Do punctuation and spaces count? Counting spaces and punctuation in reading speed.
Every language edition describes the same English-dominant NASA files and reference checks.