Lesson 7 of 8 · 6 min read

Captions, QA and packaging

Meet the WCAG rules for video, build captions from the script, check the final render, and test the exact file in your LMS before release.

What the accessibility standard asks

WCAG 2.2 is the W3C’s accessibility standard for web content, and a course page in an LMS is web content. Four of its success criteria apply to a training video. 1.2.2 Captions (Prerecorded) is Level A. 1.2.3 Audio Description or Media Alternative (Prerecorded) is also Level A.

1.2.5 Audio Description (Prerecorded) is Level AA. 1.4.2 Audio Control is Level A. Under 1.4.2, audio that plays on its own for more than 3 seconds needs a way to pause or stop it, or a separate volume control.

Inference: if the narration describes everything important on screen, the video needs little separate audio description. Lesson 3’s rule to narrate what the screen shows also serves 1.2.5. A transcript meets 1.2.3 at Level A, but not 1.2.5 at Level AA.

Leave out a music bed under the narration. Music makes the voice harder to follow, and if the video plays on its own it brings 1.4.2 into play. One learner reported modules with an AI voice over dramatic music, and could not recall the content 30 minutes later.

Captions are part of an employer’s duty

In 2019 the US Equal Employment Opportunity Commission sued Powerlink Facilities Management Services. The EEOC alleged that the company’s orientation videos had no closed captions, so a hearing-impaired employee could not follow them. It alleged that the company kept the employee from starting work for months instead of providing an accommodation.

The case shows how US law treats training videos for private employers. ADA Title I covers job training, so a deaf employee may need captions as an accommodation. Captioning every video before release is simpler than captioning it after someone asks, while that person waits.

Public bodies face a separate rule. ADA Title II covers state and local governments, not private employers. In April 2026, the Department of Justice moved the Title II web deadlines to April 26, 2027 for entities serving 50,000 or more people. Smaller entities have until April 26, 2028. The standard stays WCAG 2.1 Level AA.

Build captions from the script

Use closed captions that each learner can turn on and off. Lesson 3 kept full sentences off the screen, because printed narration on top of spoken narration hurt learning in Mayer’s redundancy studies. Burned-in subtitles put that text back for everyone. Closed captions meet 1.2.2 and let each learner choose.

Build the captions from the approved script, then check them against the final audio. One learner reported a course in which the AI voice said “genAI” three different ways. The captions had been transcribed from the audio, not taken from the script, so they repeated the errors.

Use the written form in captions, not the respelling you gave the voice. The voice reads “L-K-S”, but the caption shows “LKS”. A person reads every caption while listening, line by line.

Check three things in each caption. The words match the audio. The caption appears while the words are spoken. Names and codes use the written form from the pronunciation list.

The QA pass

Run one QA pass on the final render, with the script and the pronunciation list open. Check six things in order. Facts first: a colleague who knows the process confirms each step and each rule. Then pronunciation: listen for every term on the list, in every segment.

Then lip-sync, if the video has an avatar: watch the mouth on each sentence. Then generated shots: check every frame of every generated clip against what it should show. One learner reported a safety module in which two generated forklifts collided and their loads merged into each other.

Then the final render, not the preview. Buyers of avatar tools report previews that did not match the rendered video. Last, the disclosure: confirm that it is spoken in the first seconds and appears in the captions. Lesson 6 explains why it needs both.

If the video has a presenter, ask one more question with the final render in front of you: does the medium fit the message? Lesson 1 noted that an AI presenter on a course about AI, security or ethics costs trust.

Write down each check, the date and the name of the person who did it. When the process changes, the record shows which checks to repeat. Lesson 8 returns to updates.

Package and test in your LMS

Test the video in the LMS your learners use, on the devices they use, before release. Check that the captions turn on, that the player can pause, and that the course records what you need. EdzLMS, a learning-platform vendor, puts this test in its production checklist.

With a SCORM plan, the update path matters. Synthesia’s help center says that when you replace a video with a new version, its SCORM file updates automatically. Synthesia offers SCORM export on Enterprise plans only. Lesson 5 lists the plans that include SCORM at each vendor.

Without a SCORM plan, upload an MP4 to a course page in your LMS, or wrap it in an authoring tool that produces SCORM. One buyer reported that publishing MP4s to a video platform stripped the video’s interactive hotspots. So test the exact file you will publish, not a draft of it.

Worked example · Larkfield Supply

Riley’s QA pass finds two problems

Riley has the final renders of both segments, the approved script, the pronunciation list and the disclosure line. Morgan checks the facts. Riley does the rest with headphones on.

  1. Morgan watches both segments against the returns policy. Every step is right, including the 30-day check from the delivery date.
  2. Riley listens for each term on the pronunciation list. Segment one is correct. In segment two, the voice says “Aubriot” as “AW-bree-ot”.
    What would you do next?Segment one passed with the same voice. Why would segment two fail?Think it through, then show the answer

    Riley generated segment two in a separate file in the voice tool, and the respelling lived only in segment one’s file. Riley adds the pronunciation entry to segment two, generates it again, and listens to the whole new render.

  3. Riley turns the captions on. The editor generated them from the audio, so they show “Aubriot” as “Oh Brio” and the product code as “L K S 20417”.
  4. Riley replaces them with captions built from the approved script, using the written forms “Aubriot” and “LKS-20417”. Riley then reads every caption against the audio.That answers the question you guessed at: Guess: should a training video’s captions come from its audio track?
    What would you do next?Which checks in the six-part pass can Riley skip?Think it through, then show the answer

    Lip-sync and the frame check of generated shots. The video has no avatar, and the storyboard from Lesson 4 used real photos for its cutaways. Riley writes “not applicable” beside both, so a reviewer can see they were considered.

  5. Riley checks that the disclosure is spoken in the first seconds of each segment and appears in the captions. Both segments pass.
  6. Riley uploads the final MP4s to a hidden test course in Larkfield’s LMS and plays them on a warehouse tablet. The captions turn on, and the player pauses.

Both problems were in the final files. Riley adds one rule to the video plan: check every segment again after any new render. The video goes live in the LMS on November 17, two weeks before the new returns process starts. The one-page screenshot guide from Lesson 1 sits on the same course page.

Starter prompts

  • Build my QA checklist

    Here is my video plan: [paste the script, the pronunciation list, the list of shots marked as screen recording, photo, avatar or generated, and the disclosure line]. Write a QA checklist for the final render, in the order I should check. Include facts, every pronunciation-list term by segment, lip-sync if there is an avatar, a frame check for each generated shot, the captions against the audio, and the disclosure. Do not add facts about my process. Where my plan is missing something the checklist needs, write "missing" and say what.

  • Turn my script into captions

    Here is my approved script: [paste it]. Here is my pronunciation list, with the written form and the spoken respelling of each term: [paste it]. Turn the script into caption text, with no more than two short lines per caption. Keep every word of the script unchanged, except that each respelling becomes its written form. Do not add timings, because I will time the captions against the audio. List every line you split in an awkward place.

When not to use AI

  • Auto-generated captions as the final captions. Build them from the approved script, and have a person check them against the audio.
  • An assistant’s review in place of the QA pass. A person has to watch and listen to the final render.
Recall from memoryWhere should captions come from, and what three things do you check in each one?Answer it without looking back, then show the answer

From the approved script, then checked against the final audio. The words match the audio. The caption appears while the words are spoken. Names and codes use the written form from the pronunciation list.

Exercise · your project

Write the release checklist for your video.

  1. List the WCAG criteria that apply to your video, and how you will meet each one.
  2. Write your QA checklist in order, with your pronunciation list attached.
  3. Name the colleague who will check the facts, and book the time.
  4. Find out what file types your LMS accepts, and plan a test upload of the exact file you will publish.
  5. Add the checklist to your video plan.

Sources