YouTube auto-captions, repaired by NoaLingua

Two separate faults, and everyone hits both

The first is punctuation, or the absence of it. A YouTube auto-caption track arrives as an unbroken run of lowercase words with no full stops, no commas and no capitals. It is legible for a few seconds and exhausting over an hour, and it is markedly worse in a language you are still learning, because punctuation is doing a share of the grammatical work you cannot yet supply from context.

The second is worse and much less obvious. A caption block is cut to fit the player, not to end a thought. It breaks mid-clause, splits a verb from its object, and starts a new block whenever the previous one ran out of screen. A caption block is a timing artefact — it is not a unit of meaning, and it was never intended to be one.

Why that second fault ruins a translation

Every dual-subtitle tool has to translate something, and the obvious thing to translate is the caption block that is currently on screen. That is what taking the captions as given means, and it is what almost everything in this category does.

The result is a translation of half a sentence. Machine translation is good at sentences and unreliable on fragments, because a fragment does not carry the information the translator needs — the subject that was in the previous block, the object that is in the next one. So you get a second line that is fluent, confident and quietly wrong, in a way that is very hard to notice while you are watching. You do not see an error message. You see a sentence that makes sense and is not what was said.

It also breaks everything built on top of the line. Replay, previous, next and auto-pause all land wherever the block happened to end, which is mid-clause. A repeat that starts three words into a sentence is not a repeat of anything.

What NoaLingua does before it translates a word

It repairs both faults first, and the order matters: nothing is translated until the text is a real sentence.

Caption blocks are reassembled into whole sentences, so what gets translated is a complete thought rather than a fragment that happened to fit the screen — and so replay, previous, next and auto-pause land on real sentence boundaries. Where the auto-generated track has no punctuation, a downloadable model restores the full stops and commas before that reassembly happens.

The repair is also constrained, which is the part that makes it trustworthy rather than merely clever. The punctuation model may only re-punctuate — if it returns so much as one different word, the original stands unchanged. And a full stop is never inserted just because a speaker paused, because that would be inventing sentence boundaries rather than restoring them.

What this changes while you are actually watching

The translated line is a translation of a sentence, so it says what was said. Pressing repeat replays a whole sentence from its beginning. Auto-pause stops where a thought ends, which is the only place a pause is useful. And a word you click sits in a complete sentence, so the meaning you get back is the meaning it carries there rather than a guess made from four words.

It carries through to everything downstream, too. A transcript exported as a bilingual PDF reads as prose rather than as a column of fragments, because the sentences were rebuilt before the export existed. And a sentence you save to your deck comes back as a sentence.