Subtitle readability: CPS and line length

Reading speed is measured in characters per second: take the characters in a cue, spaces included and line breaks excluded, and divide by how long it stays on screen. Netflix's English timed-text guide allows 20 characters per second for adult programming and 17 for children's; many European broadcasters work a little slower still. A cue at 30 characters per second is technically readable by a fast reader who is not looking at anything else, which is not the situation a viewer is in.

The BBC's guidelines reach the same place from a different direction, working in words per minute — roughly 160 to 180 for pre-recorded programmes — and reminding subtitlers that the rate should follow the speech rather than a spreadsheet. Both approaches agree on the important part: when a line is too fast, you extend it or you cut words, and cutting words is usually the better answer.

Condensing is a craft, not a defeat. Spoken language is full of repetitions, false starts and names the viewer already heard; a subtitle that keeps the meaning and drops the filler reads faster and feels more accurate than a literal transcript.

Forty-two characters per line is the modern default for Latin scripts, adopted by Netflix and by the TED translation guidelines among others. The BBC works to a narrower line, in the high thirties. The exact figure comes from screen width and font size, not from typography — a line that is too long forces the eye to travel across the whole picture and back.

Two lines is the ceiling. A third line eats into the picture and, worse, pushes the eye into a scanning pattern that takes longer than reading two lines of the same total length. When a cue genuinely needs three lines, the answer is almost always to split it into two cues.

Balance the two lines rather than filling the first one to the brim. A line of 40 characters followed by a line of 6 looks broken and reads worse than 24 and 22, which is why a decent re-wrap squeezes the column to the narrowest width that still fits the text in two lines.

The usual floor is around five sixths of a second — twenty frames at 24 fps — because anything shorter registers as a flash rather than as text. The usual ceiling is seven seconds: past that, viewers start re-reading a line they have already finished, which is oddly uncomfortable to watch.

Between those limits, duration should follow the length of the text through the reading-speed figure, not the length of the spoken line. A three-word answer that takes four seconds to say does not need a four-second subtitle; it needs about one and a half, timed to the voice.

Consecutive subtitles need a gap, or the viewer cannot tell that the text changed. Two frames — about 80 ms at 24 fps — is the common minimum, and it is also the reason an automatic overlap fix should trim the earlier cue rather than push the later one forward.

Break at a syntactic joint. Keep an article with its noun, a preposition with the phrase it governs, a first name with its surname, and a verb with its auxiliary. A break that lands between an adjective and the noun it describes forces the reader to hold an incomplete idea across a line, which costs real milliseconds.

When the text contains a comma or a full stop near the middle, break there: punctuation already marks the place where the sentence takes a breath. Only when there is no such joint should you fall back to pure balancing, and even then it is worth checking that the result does not read as two half-thoughts.

Dialogue between two speakers in one cue is set as two lines, each opened with a dash. That convention exists precisely because the line break is doing grammatical work: it separates speakers, so it must not also be separating a phrase.

A subtitle that crosses a cut makes the viewer re-read it, because a change of image resets attention. Professional practice is to snap cue boundaries to shot changes when they fall within a few frames, and otherwise to keep the whole cue inside one shot.

Position matters when the picture has burned-in text. WebVTT cue settings and ASS position tags both exist to move a line to the top of the frame for exactly this reason; plain SRT cannot do it, which is one of the few situations where the format genuinely limits you.

Five checks catch nearly everything: cues faster than your reading-speed limit, cues shorter than a second or longer than seven, lines over your character limit, cues with more than two lines, and overlaps. Each one is a number you can compute from the file itself, without watching a frame of video.

SubShiftly shows those numbers beside every cue and turns them red when a limit is passed, so filtering the table to the problem rows tells you in one glance how much work is left. The cleanup pass can then re-wrap lines, trim overlaps, extend cues that are too short and drop the empty ones — and what remains red afterwards is the handful of cues that need a human to shorten the text.