A subtitle QA checklist that separates syntax from editorial review

A subtitle file can be syntactically valid and still be difficult to follow. It can also read beautifully in an editor and fail when the destination player parses it.

A dependable QA process therefore has two separate passes:

1. Machine validation checks whether the artifact follows its format and timeline rules.
2. Editorial review checks whether the synchronized text conveys the information a viewer needs.

Neither pass replaces the other.

Pass one: validate the artifact

Run validation against the exported file, not only the editor’s internal data. At minimum, check:

• the expected encoding and line endings survive export;
• required headers are present, including WEBVTT for WebVTT;
• cue timestamps parse in the selected format;
• each cue starts before it ends;
• cue order is intentional;
• blank cues and duplicate cues are reported;
• unsupported markup and cue settings are identified;
• the final timestamp is within the media duration;
• conversion did not silently remove metadata or positioning.

Do not apply one format’s grammar to another. SRT commonly uses 00:00:01,250; WebVTT uses 00:00:01.250. WebVTT also defines cue settings and parsing behavior that SRT does not share.

Overlap is a policy check rather than a universal automatic error. It may represent simultaneous speakers, but destination players handle overlapping cues differently. Flag it for review and test the target platform.

Pass two: listen and watch

Play the media with the captions visible and the sound muted for part of the review. Then listen with the image hidden or reduced. These perspectives expose different omissions.

Check that:

• speech is synchronized closely enough to identify who is talking;
• a cue does not vanish before it can be read;
• sentence and phrase boundaries feel natural;
• speaker changes are clear when the image is ambiguous;
• meaningful music, laughter, alarms, and off-screen sounds are represented;
• captions do not cover names, diagrams, demonstrations, or other relevant visuals;
• corrections preserve what was said rather than silently rewriting the speaker’s meaning.

WCAG’s explanation of captions explicitly includes speaker identification and non-speech audio needed to understand the content. This makes a dialogue-only transcript an incomplete caption track in media where sound carries meaning.

Treat reading rules as editorial policy

Character limits, characters per second, minimum duration, and line count are useful production controls, but there is no single value that fits every language, audience, viewport, and delivery platform.

Record the policy used for a project. A reviewer should be able to tell whether a warning comes from:

• the subtitle file format;
• the destination platform’s delivery specification;
• a broadcaster or client style guide;
• a language-specific editorial rule;
• a local product default that can be changed.

This distinction prevents a house rule from being presented as a WebVTT or accessibility requirement.

Test the text beyond English

A fixture with only ASCII English misses common export failures. Include representative samples for:

• accented Latin characters and combining marks;
• CJK text without space-delimited words;
• right-to-left Arabic or Hebrew with numbers and punctuation;
• emoji and characters outside the Basic Multilingual Plane;
• speaker names that mix scripts;
• vertical text if the destination and format claim to support it.

Review bidirectional text in the final player. A text editor showing the right characters does not prove that ordering, alignment, or punctuation will render correctly on video.

Verify names and terminology separately

Proper nouns, product names, technical terms, and abbreviations deserve a focused pass. Search for each known name rather than waiting to notice errors during playback.

When the recording is unclear, mark uncertainty for review instead of confidently inventing a word. Automatic confidence scores can prioritize attention, but they are model signals, not proof that a token is correct or wrong.

Re-run QA after every transformation

Edits, translation, frame-rate conversion, time shifting, and format conversion can each introduce new defects. Preserve the source revision and transformation settings so the result is reproducible.

A compact release sequence is:

1. freeze the reviewed source revision;

2. export the destination format

   ;

3. run parser and timeline checks;

4. inspect all warnings;

5. play representative opening, middle, and ending sections;

6. verify multilingual and simultaneous-speaker cases;

7. archive the validated file with its media version.

The result is not “perfect captions” certified by a tool. It is a specific artifact that passed documented technical and editorial checks for a known destination.