Captions checklist
How to check the accuracy of auto-generated captions.
Seven checks, in the order that finds errors fastest.
Play the film with the track on and check seven things in this order: names and jargon, numbers, speaker changes, sound cues, line length, reading speed and timing. The first four are about whether the words are right. The last three are about whether a person can read them in time. W3C says automatic captions do not meet accessibility requirements until someone confirms they are fully accurate, so this pass is required, not optional.
Video hosting for filmmakers · 5 GB free · Paid plans from USD 9/month
By uncompressed.io · Updated September 2026
Why the pass is required
A generated track is a draft
until someone checks it.
WCAG 2.2 makes captions for prerecorded video a Level A requirement (success criterion 1.2.2), and captions for live audio Level AA (1.2.4). W3C’s own guidance says automatic captions do not meet that requirement until they are confirmed to be fully accurate. It calls them a starting point that you will need to edit.
A
WCAG 2.2 level
1.2.2 Captions (Prerecorded)
42
Characters per line
Netflix English (USA), two lines at most
20 cps
Reading speed
Netflix adult programs; 17 for children
7 s
Longest event
Netflix; shortest is 5/6 of a second
“Accurate” covers more than the dialogue. W3C’s Understanding document for 1.2.2 says captions also identify who is speaking and include meaningful sound effects. The DCMP Captioning Key names five elements of quality captioning: accurate, consistent, clear, readable and equal. A track can get every word right and still fail because nobody can read it in time. That is why the checklist below has two halves.
The checklist
First whether the words are right,
then whether anyone can read them.
Go in this order. The first checks catch the errors a client notices. The later ones are mechanical and quick once the words are fixed. If you fix the words last, you will have to check the line breaks and timing again.
Names and jargon
Make a list before you press play: people, the company, products, places, acronyms and trade terms. Search for each one, and for the words it most likely turned into, in the exported .srt in a text editor or in the review room's dialogue search. A misheard name can recur, so check every occurrence, not just the first. Check spellings against the lower thirds, the brief or the client's website, not your memory.
Numbers
Check every figure against the audio: amounts, percentages, dates, years, phone numbers, model numbers. Similar-sounding numbers are easy to mix up, and so are figures where the speaker drops a word, like "one fifty" for 150 or 1.50. Choose one style for writing numbers and use it throughout. Consistency is one of DCMP's five elements.
Speaker changes
Wherever the speaker changes, check that the new line starts a new caption and does not continue the last speaker's sentence. If the viewer cannot tell who is talking, because the speaker is off screen, speaking in voice-over or talking over someone else, add a speaker label. W3C lists identifying the speaker as part of what captions do.
Sound cues
Listen for sound that affects the meaning and has no text: a phone ringing, laughter, music starting or stopping, a sound the speaker reacts to. W3C's Understanding document includes meaningful sound effects in captions, and its caption guidance shows an automatic track getting words wrong in ways that change the meaning. Leave out background sound that does not change the meaning.
Line length
Keep lines short enough to read at a glance. Netflix's English (USA) guide allows 42 characters per line and two lines at most. Break lines between phrases, not in the middle of one: keep an article with its noun and a name on one line.
Reading speed
Divide each caption's characters by the seconds it is on screen. Netflix allows up to 20 characters per second for adult programs and 17 for children's programs. DCMP measures in words per minute: 130, 140 or 160 for educational media depending on level, with speaker IDs and sound effects counted as words. If a caption is too fast, shorten the wording or split it across two events.
Timing
Each caption should appear with its words and leave when they end. Look for captions that start early, stay up through a cut or cover a pause. Netflix's general requirements set a minimum of five sixths of a second per event (20 frames at 24 fps) and a maximum of 7 seconds.
The arithmetic that catches most reading-speed problems
The published limits, side by side
Two style guides, two ways of counting,
and none of them is the only one.
Netflix figures come from its English (USA) Timed Text Style Guide and its general requirements. DCMP figures come from the Captioning Key pages on quality and presentation rate, and the reading rates apply to educational media. Both were fetched 17 September 2026. If your contract or distributor sends a spec, follow that spec.
Doing the pass here
You edit the words,
the client submits the spellings.
Captions here begin as a draft from a speech-to-text service. They are generated only when someone asks, and the minutes come out of a monthly allowance set by your plan. The free Starter plan includes 20 minutes. In the owner’s editor, the transcript follows playback, and clicking a line or timestamp jumps the player there. That makes the timing check a matter of clicking and watching. You edit the text inline and save, then export the track as .srt or .vtt. The editor is where you fix the words. If a cue needs new start or end times, adjust them in a subtitle tool after export; this editor changes text, not timing.
The first check on the list, names and jargon, is the one where your client knows more than you do. In the review room, a reviewer on the share link can read the transcript, search the dialogue for a word and jump to that line. A reviewer cannot change the captions. They submit a correction, and you apply it. This is on by default, and you can switch it off for any film. Saving new caption text clears any earlier approvals tied to those captions, so a client who signed off on the old wording signs off again.
You do
Numbers, speaker changes, sound cues, line breaks, reading speed and timing. These checks depend on the picture and the edit, which you know better than anyone.
The client submits
Spellings of their own staff, products and acronyms. They send a correction on the track instead of a list in an email.
You apply
Each correction you accept, then save. The file you export is the corrected track.
When this is not enough
If a standard is in the contract,
hire the specialist.
This checklist is a working pass for corporate, commercial and social work where no captioning standard is written into the contract. It does not replace a professional captioning service when a contract, broadcaster or distributor names a standard and checks the delivery against it. In that case, send the spec to a captioning service, budget for it, and use the generated track as a working transcript. For the wider workflow, see captions for client video work. To decide between a sidecar file and burned-in text, see burned-in subtitles or an .srt file.
Questions
Frequently asked
How accurate do captions need to be?
The published guidance sets no error rate you can aim for and stop. W3C says automatic captions do not meet user needs or accessibility requirements unless they are confirmed to be fully accurate, and the DCMP Captioning Key says errorless captions are the goal. Plan to read every line.
Can I check captions without watching the whole film?
Only partly. You can find names, numbers and jargon by searching the text, in the exported file or in the review room. You cannot check a missing sound cue, a speaker change or a caption that stays up through a cut without watching the picture. Search first, then watch the whole film once with the track on.
Which reading speed should I use?
Use whatever your delivery spec names. Without one, two published figures are reasonable references. Netflix's English (USA) guide allows up to 20 characters per second for adult programs and 17 for children's programs. The DCMP Captioning Key sets educational media at no more than 130, 140 or 160 words per minute depending on level, and counts speaker IDs and sound effects as words.
Can a client edit the captions in the review room?
No. A reviewer on the share link can read the transcript, search it and submit a correction. The owner decides whether to apply it. Reviewer caption access is on by default, and the owner can switch it off for any film.
Let the client fix their own product names
Generate the draft, do the pass, and let the reviewer submit corrections for the names only they know. You apply them. Caption generation is metered in minutes a month by plan, and the free Starter plan includes 20.
Sources
- 1.W3C, Web Content Accessibility Guidelines (WCAG) 2.2, Recommendation of 12 December 2024: 1.2.2 Captions (Prerecorded), Level A, and 1.2.4 Captions (Live), Level AA (fetched 17 September 2026)
- 2.W3C, Understanding Success Criterion 1.2.2: captions identify who is speaking and include meaningful sound effects (fetched 17 September 2026)
- 3.W3C Web Accessibility Initiative, Captions/Subtitles: automatic captions need editing and are a starting point (fetched 17 September 2026)
- 4.DCMP Captioning Key: elements of quality captioning (accurate, consistent, clear, readable, equal) (fetched 17 September 2026)
- 5.DCMP Captioning Key: presentation rate for educational media (130, 140 and 160 words per minute) (fetched 17 September 2026)
- 6.Netflix Partner Help Center, English (USA) Timed Text Style Guide: 42 characters per line, two lines, 20 and 17 characters per second (fetched 17 September 2026)
- 7.Netflix Partner Help Center, Timed Text Style Guide General Requirements: five sixths of a second minimum and 7 seconds maximum per subtitle event (fetched 17 September 2026)