changelog

// how it got built, and what each round of it fixed

this show is the second one through a pipeline that was built for the first, and most of what[about] describes was already working before an episode of it had been transcribed. what is here is the start of a back catalogue eleven years long, and the record of what it took to make the first pass worth reading.

[where it stands]

counted off the correction files themselves, so it moves whenever another round is worked. worth reading against what it is aimed at — 10 episodes, 14 hours, 151,000 words of English — since the whole point of what follows is that almost none of that has to be looked at:

85

terms in the show-wide dictionary, 3 of them forced rewrites

35

terms pinned to a fixed German and French rendering

[before the first episode]

a transcriber that has never heard of Ozone, X-Alps orGuschlbauer spells all three the way they sound, and every later correction is then a guess made without the audio. the usual way to build that vocabulary is to read the transcripts and see what went wrong — which is circular when there are none.

read the publisher's own pages

every episode of this show has a page on the publisher's site, with notes, chapter lists and a "mentioned in this episode" list of names. all 278 were fetched and parsed, and the words that recur across many of them — rather than the ones that appear once, which is what a mishearing looks like — became the show's first dictionary. it is the same signal the other show's checks use, read with the opposite sign.

and take the publisher's typos with a pinch

the notes spell Paul Guschlbauer two ways across 38 pages. so a proposed term is reported together with any near-spelling of it in the same source, because putting the commoner of two into the dictionary would make this pipeline repeat somebody else's typo in every episode.

stop assuming German

the phonetic matching that repairs a misheard name was written for German, where v andw sound alike and a final d and t are the same sound. applied to English it merges van with wan and bad with bat, so the rules that are wrong for English were dropped and th stopped folding tot. those are a judgement about English spelling, not a measurement — the first episodes are what will settle them.

[what is still wrong]

plenty, and less is known about it here than on the other show. the checks that find a mistake without anyone reading the whole thing need a corpus to compare against, and this one is a few episodes deep — so for now the dictionary is doing most of the work and it was built from somebody else's writing rather than from this pipeline's own mistakes. a confidently wrong sentence reads as perfectly good English and goes past everything. when a line matters, listen to it: it is one click away, which is the only real answer this site has.

← about · all episodes