The Rule This Lesson Has to Answer
The previous lesson made a strong claim: the second language is written, not translated, re-authored from the picture in its own rhythm, and phrase-for-phrase correspondence proves it was not. On 2026-10-04 the workshop met a job where that rule cannot be followed as written, and how it bent is the useful part.
The channel's owner asked for his published talks to be made again in English, Chinese, and Japanese for Pippa's own channel, instead of leaving foreign viewers to the platform's automatic dubbing and captions, which translate without knowing anything about the talk. His design had four parts: Pippa translates knowing the whole conversation; a Pippa who has never seen the original makes the result read natively; the voices are new in each language; and each edition goes up private to her channel until he publishes it.
Two Sources, Both Needed
The pointer for this pipeline is a new kind: the published video's address. The conversations it was cut from ride along as sources. Both are needed, and the reason is precise. The published film is what the viewer actually saw, including every cut the owner made in his own editor. The conversation is the ground truth those cuts were made from, with the context a translator needs. A translation built from the conversation alone would restore what he removed; one built from the video alone would translate without context.
So the first stage locates one inside the other. The family's transcription engine already holds the published video's timed transcript; the pipeline compares it with the candidate talks and keeps each side of the conversation that the video demonstrably contains, giving a one-word reply its neighbour's verdict and listing every side the video holds only partly, and every stretch of the video no turn accounts for, for the producer to judge with him. On the first film it found both talks the video joins, kept the broadcast's turns, dropped the after-show, and found the one passage he had cut out because a turn had gone wrong.
The First Method Was Rejected
The first English edition re-made the film from the conversation: a translated script, new pictures, new subtitles, both voices synthesized. The owner rejected it on sight. It looked machine-made, and he named why: his master carries the effects he put in by hand, and only the master can carry them. A re-made film is a new film that happens to share a script.
The method became the dub. His master is used as it is. Only the voices change, sentence by sentence, and only Pippa's own on-air cards are redrawn in the new language — found by their navy box and by the moment the pixels in their slot change, never by searching the screen for text. The original pipeline's re-make path still exists in the code. It is no longer the method.
Translating Under a Fixed Picture
This is where the previous lesson's rule meets its boundary. Re-authoring from the picture assumes the author controls the picture's timing. Here the picture is fixed, and every sentence has a room: the time until the next Korean sentence begins. So the dub separates the master into voice and background once per video, tells his voice from hers by pitch, and gives each of the transcript's timed sentences to its turn. The producer's sheet lists, for each timed sentence, the side of the conversation it belongs to and who said it, its room, the Korean, and the line to say; the pitch reading of every sentence the separated voice lets it measure sits beside them as its own field, and it steers the matching itself: placing a sentence on the other speaker's side costs it a fixed penalty in the matching score, which is what keeps her short openings off his turns (a penalty, not a wall: strong enough evidence can still outweigh it). Each line is read through the voice engine with its neighbouring lines as context and laid where its Korean sentence began, or just after the line before it when that one ran long, sped up by at most fifteen percent, and mixed over the original background with the speakers' laughs kept from the original voices. A line that still does not fit goes on an overflow list instead of being squeezed, and because every overflow pushes the lines after it late, that list, longest first, is the list of lines to shorten. And his loose way of talking is kept on purpose: it does not need to be perfect, only natural.
The guard against translated-sounding text did not disappear; it moved. The rule's purpose — the second audience should receive what the first one did, in language that does not read as converted — is now held by a reader instead of a writing method. A separate brain reads the finished translation with the source withheld (each line with its speaker and its time budget, the title and synopsis, and every piece of translated screen text, but never the Korean), files graded findings, and the author applies or rejects each one with a written reason. A reviewer who never saw the Korean cannot be soothed by its familiar shape, so phrase-for-phrase stiffness is exactly what it notices.
Each Language Is Its Own Delivery
Each edition is rendered over the master at about its original bitrate, with the redrawn cards and a seven-second badge saying Pippa made it alone, and uploaded once, private, to Pippa's channel. Then one finishing step binds every language into the production's single record: each render's hash, its measured duration, streams, and loudness, every card check and placement list. Fonts are chosen per language — Japanese kanji forms in a Japanese face, Chinese in a Chinese one — because the card renderer passes the text no language tag, and without one a face draws the shared characters in the shapes of the language it was made for. A face that holds both sets of forms, given text tagged with its language, would also work; this renderer chose two faces. And a Japanese range mark that the face lacked once had to be routed to a face that has it.