What we learned
building this
Every post here comes from this project's own measurements — the failures that shipped, the checks that caught them, and what a season actually costs. That is the one thing we can write that a competitor cannot copy.
Why sound labels come from a table, not from the model
[LAUGHS], [DOOR CLOSES], [PHONE RINGS]. Subtitles for the deaf and hard of hearing are full of them, and a language model gets them subtly wrong in ways no reader can catch. One table per language, and the Arabic bug that settled the design.
Read →How we score a machine translation against a human subtitler
Every prompt change here is measured against human-translated subtitles before it ships. How the reference is aligned, what chrF++ does and does not tell you, how big the noise floor is, and why the numbers cannot be compared across languages.
Read →Translating subtitles into Bulgarian: what the alphabet does not tell you
Bulgarian was this pipeline's first language and the source of its first regression test. Five things that go wrong when English subtitles are translated into Bulgarian, each with the check that now catches it.
Read →The 24 characters per second rule, and what translation does to it
Reading speed belongs to the viewer, not the language. Measured across eight languages: the English source already breaks the limit more often than any translation does, and most residual violations cannot be fixed without deleting dialogue.
Read →How to translate an SRT file without losing a line
An SRT file is four things per cue and every one of them can be damaged by a translator. What the format actually is, where translation breaks it, and a five-minute check you can run on any translated file.
Read →What it actually costs to translate a season of subtitles
Measured, not estimated: real API cost per episode, why the second person to order a popular title pays the same but costs us a third, and what that means for pricing.
Read →What the checks on a subtitle cue actually catch
Every cue is validated before it reaches you, and anything that fails goes back to the model with the complaint attached. Here is what the checks look for and why each one exists.
Read →Why free subtitle translators quietly lose half your dialogue
A subtitle cue can contain a newline. A lot of translation pipelines pass cues around in a format where a newline ends the record. Here is what that does to an episode.
Read →