How we choose subtitle cue boundaries
Cue boundaries are a readability decision, not just a timing one. Here is the rule set we use to split a word-aligned transcript into readable lines.
Word-level timings tell you when each word was spoken. They do not tell you where a subtitle line should end. That second question is the one viewers actually notice, and it is answered by rules, not by the waveform.
The constraint is reading speed, not audio
A viewer reads a subtitle while also watching the picture. The industry working limit is around 17 characters per second for adults watching in a second language, and 20 for comfortable native viewing. A cue that is on screen for two seconds can therefore hold roughly 34–40 characters — about 12–14 characters per line across two lines, which is why most style guides cap lines at 42 characters and prefer two lines.
Timings that ignore this produce cues that are technically accurate and unreadable: the words were said then, but nobody could have read them then.
The rules, in priority order
- Never split inside a grammatical unit. A noun phrase, a prepositional phrase, a subject from its verb. If the audio forces a boundary there, prefer extending the previous cue over cutting mid-phrase.
- Break at the strongest syntactic boundary available. Clause boundary beats conjunction beats comma beats nothing.
- Balance the two lines. A 30/10 split reads worse than 20/20, because the eye spends the first line accelerating into a short second line.
- Keep the last line a whole thought. A cue ending on “the” or “of” makes the viewer’s eye do work the next cue immediately undoes.
- Hold, don’t flash. A cue under ~0.8s is a subliminal flash. Merge it with a neighbour even if the merge is slightly longer than the speech.
Rules 1 and 5 conflict constantly, which is why cueing resists full automation. The products that claim otherwise are usually applying rule 5 alone.
Where forced alignment helps
Because the aligner knows the timestamp of every word, not just the sentence, it can evaluate a candidate break against the real audio. “Split after this word” is a question with a measurable answer: does this word end at a place where the next starts more than 200ms later? If yes, that is a natural pause and the best break available.
Transcription-based tools cannot do this, because their word timings inherit the recogniser’s mistakes: a misheard word has a plausible-looking timestamp attached to the wrong string, and a cue split placed on it lands somewhere the speaker never paused.
What we do not do
We do not auto-shorten text. If a line cannot fit in its duration at a readable speed, the honest output is a warning, not a silently truncated subtitle. Truncating changes what the speaker said — which is a different product than the one being sold here.