Mending the transcript first

A pasted transcript is what YouTube heard, cut where YouTube chose to cut it — and one made by speech to text is what Whisper heard, cut where Whisper chose to. Both are often wrong: a name misheard, a sentence split across two captions, a caption that begins three seconds late, two words glued to the caption before. Every road out of the add page — the prompt, the blank draft, the checker — reads the panel as it stands, so a panel that is wrong makes a video that is wrong.

So the add page has a small subtitle editor, in the manner of Subtitle Edit, for the one moment it is wanted: before anything is built from the transcript. Paste the transcript (step 2 of Adding a video), then press Edit the transcript… under the box.

1. The list#

The editor, The transcript, says how many captions it read and lists them, one row each:

A caption runs until the next one begins, so moving one moves where the one before it ends; there is no end to set.

After ✂, a transcript Parseh made with word times starts the second half at its first recorded word. A pasted transcript still uses the former halfway guess and says so. A join puts the two captions’ words together with a space; if it would remove a start you set by hand, it asks before doing that.

Above the list:

shift

n s, move them all — add that many seconds (negative for earlier) to every caption’s start at once. A panel copied from a re-uploaded video is often late or early by the same amount throughout. It refuses to take the first caption before the video begins.

+ caption at the end

— one more, four seconds after the last, to give its words and its time.

✨ tidy up

, its text cues / recorded pauses choice, undo, redo and or with an LLM… — below.

Starts from the recording#

When Parseh made the transcript, it carries the heard words behind the panel while this editor is open. A cut, tidy result and accepted LLM answer takes a new caption’s start from its first matched word. A start you type or nudge is kept as a person-set boundary; automatic work cannot quietly discard it. The editor’s count explains how many starts are from the recording and how many are worked out. This timing detail belongs only to a Parseh-made transcript: a pasted transcript remains the familiar panel editor.

Hearing a nudge#

For a YouTube video, the video itself stands over the list, and ▶ plays one caption and stops where the next begins. That is how a nudge is answered: press, play, press again. If the video cannot load (no internet), its place says the video did not load — the times can still be typed.

A film on this machine is not in the player yet while it is being added, so there is no video over the list: its captions are timed by eye and by typing. The ▶ in each row stays, but cannot play a film: it answers the video is still loading.

A tenth of a second#

The panel keeps what the steps make. A caption’s clock line takes a fraction — 0:08.4 — written only where there is one, so a panel of whole seconds goes back as whole seconds, and every transcript written before reads exactly as it did. The player seeks to a tenth as readily as to a whole second. A pasted .srt keeps its cues’ own fractions.

2. ✨ tidy up#

An automatic transcript is cut where YouTube ran out of room, never where a sentence ends. ✨ tidy up reads the whole transcript at once and re-cuts it into sentences, one caption each:

1. With a Parseh-made transcript, it reads the carried word times. With a pasted transcript it lays words across the time their caption covers as before (a language written without spaces, character by character).
2. It cuts that stream into sentences — across the caption edges, which is what the transcript cannot do for itself.
3. Each sentence becomes a caption starting at its first word’s time, and ends with the stop its language prints: . and ?, ؟ in the Arabic script, 。 and ？ for Japanese and Chinese, । for Devanagari — a question mark where one of the language’s own question words opens or closes it.

Every language here gives it something to read:

the stops the transcript already carries

— YouTube punctuates its automatic captions for English, Italian, French, German, Spanish and Japanese, and for those this is nearly the whole job;

the verb

, where the language puts it last and nothing is punctuated — Persian, Turkish: the toolbox knows the copulas from the language’s own word lists and the verbs from the installed dictionary;

a pause

the caption edges betray, a question asked without a verb at a caption’s edge, and a [music] tag, which is kept on a line of its own — these belong to every language.

Not one word changes: only the punctuation is new, and a tidy that could not promise that gives the captions back untouched. It is a guess, so it is a button and not a rule: the editor says what it did (40 captions became 23), undo the tidy takes it back in one press, and whatever it got wrong is a ✂ or a ⤓ away.

What decides whether it is offered is the dictionary of the video’s language. Not because a full stop needs one, but because everything above a full stop is a question about a word — is it a verb, is it a name — and a dictionary is what answers that. Without it the button is not shown at all; install the dictionary from the reading-help page (Settings → Reading help) and it is there the next time you open the editor, with nothing to restart.

3. Or with an LLM…#

or with an LLM… is the other road — the one the glossing step already walks. It copies to the clipboard a prompt that sets out the same job — one caption per sentence, keep the words, punctuate, time each caption at its first word, mend only a word that is plainly misheard — with the transcript inside it, and opens a box for the answer. Paste the model’s whole answer there and press use the answer; it is read as an ordinary transcript. copy the prompt again copies it once more, and close folds the box away.

The two roads fail differently, which is why both are offered: the algorithm cannot hear a misheard word and will never invent one; a model hears the sense and may invent anything. So when a model’s answer goes in, the editor says how far the letters moved — 40 captions became 23, and not one letter of the words changed, or — and the words are not the same: 3 letters more or fewer. Look them over. undo the tidy takes an answer back too. The prompt needs nothing installed.

Beside the button that copies the prompt is a menu, prompt: Parseh’s ▾: Parseh’s own prompt for this job, and the ones you wrote for it (Your own prompts). Choose one and it is what is copied, and the line it opens with says so (… · custom: keep the captions). A prompt of yours can add a rule after Parseh’s — never correct a name, leave out the tags — or stand in place of Parseh’s rules; either way the answer’s shape, one fenced block in the panel’s own form, and the transcript are Parseh’s, so what comes back goes into the box as an ordinary panel whichever prompt asked.

4. Putting it back#

Nothing is written until Use this transcript: the captions go back into the transcript box as an ordinary panel, which is what the next step reads, and the page says the transcript was edited (or unchanged). The panel comes back in its plainest form — a chapter marker as Chapter 1: …, no spoken-duration lines — which reads exactly as before. A prompt you had already prepared is forgotten, since it was made from the old panel.

It refuses, by name, a caption whose start does not come after the one before it — caption 7 starts at 0:21, which is not after the one before it (0:22) — and marks that row: the checker would refuse it later, and later is worse. It also refuses a transcript whose every caption is empty.

Cancel, ✕ or Esc leave the box as it was (left as it was).

A transcript made by speech to text is edited by this same editor, and Whisper is nowhere in it: the editor opens on what is in the box, whoever put it there, and gives the box back a panel. A generated transcript is always yours to correct before the video is added — a misheard name, a caption cut in the wrong place, a start a tenth late — and what you write over it is what the video is made of. The block that made it does not touch the box again unless you press Transcribe again, and then it asks first. While a YouTube video is being recorded there, the editor says it must wait: it would play a second copy of the same video.

5. Only while the video is being added#

There is no such editor in the player, and no speech to text either: a video already in the library has no transcribe, no transcribe again and no Whisper model to choose in its player — what a person adds is corrected with the tools every transcript has. Once a video is added, its transcript is what its glosses were checked against, and a door that could rewrite what a caption says would quietly unmake that check. The one thing that may still move afterwards is where a caption starts — the timings — and a phrase whose words YouTube got wrong has its own way out (Editing a phrase).