Subtitle to Text

Taking the timings out is the easy half. Anyone can delete every line with an arrow in it and be left with the words, and the result reads badly for three reasons that have nothing to do with the words themselves.

A cue is not a sentence. A subtitle file breaks a line wherever it ran out of screen or wherever the speech paused, so a caption transcript made by putting one cue per line is a page of fragments — half of them starting mid-clause. Joining them back up means deciding where a paragraph genuinely ends, and there are only two honest signals for that: the previous cue finished on a full stop, or the silence before this one ran past three seconds. Both are used, and nothing else is guessed.

Rolling captions repeat themselves. Automatic captions on YouTube and in most live systems scroll: each cue reprints the line above it and adds one underneath, so the same sentence appears in two or three consecutive cues. Concatenate them and the transcript says everything two or three times. That is a property of how the format displays, not a fault in the transcription, and it is why a naive extract text from SRT job produces a document three times too long. Consecutive identical lines are collapsed to one.

The markup was never meant to be read. Italics, positioning tags and WebVTT cue settings are instructions to a renderer. In a transcript they are noise, so they are removed — which also means the word count you get back is a word count of the speech.

What is deliberately not done: no punctuation is added, no capitalisation is corrected, no speaker is inferred. Every one of those would be this tool writing something nobody said.

Press Load a sample. It is four cues of a rolling caption — the first three are one sentence delivered the way an automatic captioner delivers it, and the fourth is a new scene half a minute later.

Count the lines in the input: seven, across four cues. The status line reads lines 7. Now read the output — it is two paragraphs, and the first is "It has been a long time since anyone asked me that particular question." Five of those seven lines were repeats.

The paragraph break falls before "The following morning." and not anywhere inside the first sentence, because the gap between cue 3 and cue 4 is twenty-four seconds and every other gap is zero. The italics on that last line are gone.

Turn join the cues back into paragraphs off. Now it is four lines, one per cue, with the repeats still collapsed — the shape to use when the transcript has to keep matching the file it came from, for instance when you are marking up which cues need work.

Turn on keep the start time of each block and each paragraph gets [00:00:01] in front of it. That is the version to paste into a review document, because a note is useless without a way back to the moment it is about.

If the file is an SDH track, the sound descriptions and speaker labels will come through as words — strip them first if the transcript is meant to read as prose. And if what you actually want is the file in another subtitle format rather than as text, the converter is the page for that.