Line breaks that follow the phrase

Why auto captions break in odd places

A counting rule only knows how many words or letters are on screen. Give it 'so what you want to do is open the modifier panel' with a five-word limit and it shows 'so what you want to' and then 'do is open the modifier', which nobody would write by hand. Pauses help a little, but people rarely stop talking exactly where the grammar turns.

Bad breaks cost more on short-form video than anywhere else. Each caption is on screen for a second or two, so a viewer who has to wait for the next caption to finish a thought is a viewer who stops reading.

How Cuecaps picks the break

After transcription, a small caption model reads the words and gives each one a break score. The break picker then chooses the highest-scoring places to end captions while keeping inside your limits for words per caption, letters per caption and pauses.

• It runs in your browser: the 11.7 MB model file loads from cuecaps.com, and your words are not sent anywhere for this step.
• It labels about 1,500 words in a quarter of a second on a desktop, so it finishes long before you have picked a style.
• On hand-marked real clips it scored 0.87 F1 for breaks, against 0.52 for the word-count rules it replaced (measured with captions of up to 6 words and 32 letters).

F1 combines two things: how many of the model's breaks were right, and how many of the right breaks it found. A score of 1.0 would match the hand-marked breaks exactly, so 0.87 means the large majority of captions now end where a person would end them.

Setting caption length, then fixing the odd one

1. Drop your clip into the studio and run Find words, on your device or in the cloud.
2. In the Layout tab, set Caption length to Phrases, then choose Most words per caption, Most letters per caption and Lines (One line or Up to two). The breaks are picked again whenever you change these.
3. Play it through. If one caption still ends badly, select the word it should end on and press B, Start a new caption after this word.
4. Export the MP4, or download SRT and VTT files: their cues use the same breaks you see in the preview.

Two-line styles show break quality most, because the reader sees the whole phrase at once. These three suit longer phrases:

Transcripts in other languages

The model knows English only. Transcripts in other languages keep the rule-based breaks, which end captions on word and letter limits, pauses and sentence ends and work in any language. Manual breaks with the B key work in every language too.

Questions