Subtitle Translation QA Checklist: Reading Speed, Line Length, and Segmentation

Translation Accuracy Is Not Enough

A subtitle can be accurate sentence by sentence and still fail on screen. Viewers must read it in a fixed interval while following speech, faces, action, and on-screen text. Subtitle QA therefore has at least four independent dimensions:

1. file structure;
2. synchronization;
3. linguistic accuracy and consistency;
4. on-screen readability.

Use a file-aware SRT Translator to protect cue structure, then audit the target track with the video. The key non-obvious conclusion is that a single universal characters-per-second or characters-per-line limit does not exist. Delivery platform, language, audience, genre, and subtitle type determine the applicable rule. Pick the destination's style guide before measuring compliance.

The 10-Minute Preflight

Before the full review, run this blocker-first pass:

• The file opens in the target player or editor.
• Source and target cue counts reconcile.
• Cue IDs and timestamps remain ordered.
• The file uses the required encoding, normally UTF-8.
• No source-language cue is unintentionally blank or untranslated.
• No target cue contains an instruction, Markdown fence, or model commentary.
• The first, middle, and last minute sync with the exact video cut.
• Names, dates, prices, measurements, and URLs match the source evidence.

If any blocker fails, stop. Do not spend an hour polishing line breaks in a track aligned to the wrong video.

Download the synthetic SRT translation stress test for a reproducible preflight. It exercises formatting tags, fast dialogue, speaker labels, numeric content, and multiple scripts without using customer data.

Cue Count and Timing Integrity

When the assignment says “translate without retiming,” compare these values programmatically or in a subtitle editor:

Also check for negative duration, accidental overlaps, cues beyond the video runtime, and large unexplained gaps. An overlap may be intentional for positioning in richer subtitle formats, but plain SRT workflows should not create one accidentally.

Characters per Second and Reading Speed

Calculate visible characters per second as:

CPS = visible character count / cue duration in seconds

Exclude markup tags from the visible character count. Follow the delivery specification for whether spaces, punctuation, full-width CJK characters, and line breaks count.

Do not invent a threshold because a blog says “17 CPS is always correct.” A useful QA report includes:

• the style guide and language being enforced;
• the calculation convention;
• the configured threshold;
• the worst cues;
• whether each cue should be condensed, extended, split, or retimed.

High CPS is a review trigger, not an automatic instruction to delete meaning.

Characters per Line and Two-Line Limits

Count the rendered line, not the source-code line. Font, player width, safe area, and script all affect wrapping.

Netflix's current English timed-text guide, for example, states a 42-character maximum per line for that specific delivery context (Netflix Partner Help). That number is evidence for Netflix English delivery—not a universal law for YouTube, broadcast television, CJK subtitles, vertical video, or another client's guide.

Audit:

• maximum lines per cue;
• line length under the destination's counting rules;
• whether the player wraps a supposedly one-line cue;
• whether subtitles collide with lower thirds or UI;
• whether a bilingual cue occupies too much vertical space.

Sentence Segmentation and Natural Breaks

A good line break preserves a meaningful phrase. Avoid separating:

• an article from its noun;
• a preposition from its object;
• a pronoun from its verb;
• an auxiliary from its main verb;
• a person's first and last name;
• a number from its unit;
• a negation from the word it changes.

Bad:

The board approved the
new accessibility policy.

Better:

The board approved
the new accessibility policy.

Across cues, do not reveal a punchline, answer, or identity before the audio does. Grammatical target-language order may require paraphrasing rather than moving a crucial word into an earlier cue.

Speaker Labels, Sound Effects, and On-Screen Text

Determine whether the track is dialogue-only, subtitles for the deaf and hard of hearing, or another accessibility deliverable. The required treatment differs.

Check that:

• off-screen or ambiguous speakers are identified according to the style guide;
• meaningful sound effects are translated, not transliterated blindly;
• music cues use a consistent convention;
• labels do not consume most of a short cue;
• visible signs, messages, or UI text are included when required;
• dialogue and on-screen text do not compete unreadably.

YouTube's guidance gives examples such as [applause] and [thunder] for meaningful sounds (YouTube Help). Use the destination's convention, not a random mixture of brackets, parentheses, and speaker dashes.

Numbers, Dates, Names, and Terminology

Run a deterministic pass before judging style:

• compare every digit-bearing cue to the source;
• normalize decimal and thousands separators only when the locale requires it;
• confirm whether currencies should be converted or merely formatted;
• verify dates that could reverse month and day;
• search for each approved proper noun and forbidden variant;
• check acronyms, product UI labels, and repeated calls to action;
• confirm that units were neither translated as words nor silently converted.

A fluent wrong number is a blocker. A clumsy but correct sentence can be edited; an incorrect dosage, price, date, or legal threshold changes the content.

Spot-Check the Subtitles Against Audio

Automated QA should select high-risk cues, but a human still needs to watch them. Include:

• the fastest CPS outliers;
• the longest lines;
• the shortest-duration cues;
• every cue containing a number;
• every speaker change in a dense conversation;
• each formatting tag pattern;
• the start and end of each scene or chapter;
• a random sample from the remaining cues.

For a 120-cue file, one defensible sample is all deterministic risk cues plus the first, middle, and last cue and a random sample from every remaining 20-cue block. That is a sampling design, not proof that unchecked cues are correct.

Blocker vs Cosmetic Subtitle Errors

Do not average these into one reassuring score. One blocker should fail the release even if 119 other cues are perfect.

Downloadable QA Checklist

Copy this into the delivery ticket:

[ ] Exact video version and runtime confirmed
[ ] Destination platform, language, and style guide recorded
[ ] Source and target cue counts reconciled
[ ] IDs, timestamps, ordering, encoding, and tags validated
[ ] No missing, duplicated, or unintentionally untranslated cues
[ ] CPS and line length calculated using documented rules
[ ] Natural line and cue segmentation reviewed
[ ] Speakers, sound effects, music, and on-screen text handled consistently
[ ] Names, numbers, dates, units, URLs, and terminology verified
[ ] Risk-based sample watched at normal speed
[ ] Full target track watched or review responsibility assigned
[ ] Zero blockers; major issues resolved

For a locked-timing production workflow, pair this checklist with How to Translate an SRT File Without Changing the Timestamps. For YouTube delivery, use the YouTube subtitle upload workflow.

When Not to Use a Generic Checklist

Do not use this checklist as the only acceptance standard for regulated media, broadcast delivery, theatrical subtitles, live captions, karaoke, or accessibility contracts with explicit technical specifications. Those deliverables may define frame rates, shot-change rules, positioning, colors, speaker identifiers, reading-speed calculations, or file formats that SRT cannot represent.

In that case, this is the preflight. The contract and destination style guide are the authority.

FAQ

What is a good subtitle reading speed?

The applicable threshold depends on platform, language, audience, and style guide. Record the rule you are enforcing instead of borrowing one universal CPS number.

How many characters should a subtitle line contain?

Use the destination's specification and test the actual player. A limit published for one platform and language should not be generalized to every delivery.

Should translated subtitles keep the source line breaks?

Not necessarily. Keep timing when required, but move line breaks to natural phrase boundaries in the target language.

Is matching cue count enough to prove an SRT is valid?

No. IDs may be duplicated, timestamps malformed, text attached to the wrong cue, encoding broken, or reading speed unacceptable.

Does every translated subtitle file need a full watch-through?

For publication, a full watch-through is the strongest check. When scale forces sampling, inspect all deterministic risk cues and document what was not watched.