Translating subtitles into Bulgarian: what the alphabet does not tell you

Bulgarian ships in Cyrillic, which makes one failure trivially visible: an English line left in a Bulgarian file stands out on sight. That is the only easy part. Everything else about translating subtitles into Bulgarian is a way for output to look right and be wrong, and each of these shipped at least once before it became a check.

1. Russian words spelled entirely in Bulgarian letters

медленно, глубоко, сейчас. Every letter in those words exists in the Bulgarian alphabet. A character-set test passes them cleanly, and a large model trained on far more Russian than Bulgarian reaches for them under pressure. The only defence is a wordlist of Russianisms, grown from real files. The Bulgarian profile carries 47 of them plus 63 Bulgarian function words the translation is required to contain, so a cue that is Cyrillic but not Bulgarian is flagged rather than shipped.

2. ти and вие

English has one you. Bulgarian has two, and the choice is a statement about the relationship between two characters that has to stay consistent for a whole season. The model is told to infer it from surrounding cues and hold it for a given pair of speakers, and cues travel in batches of 32 so it can see the conversation, not one line at a time. A planned dossier of who addresses whom was measured against real output and found unnecessary: across an episode, zero cues mixed forms of address.

3. Quotation marks

Bulgarian uses „…“. A translator that emits "…" is not wrong exactly, but it reads as machine output to a Bulgarian viewer in the first minute. The style instruction names the marks explicitly, and the dangling-sentence check counts the closing “ as terminal punctuation so a correctly quoted line is not flagged as unfinished.

4. Sound labels

[LAUGHS], [DOOR CLOSES], [PHONE RINGS]. Bulgarian subtitling renders these as short verb phrases, [СЕ СМЕЕ], [ВРАТАТА СЕ ЗАТВАРЯ], [ТЕЛЕФОНЪТ ЗВЪНИ], and the model does not get to write them. They resolve from a fixed table before the text is sent, so a whole class of confident nonsense never has a chance to appear. A label the table does not know is reported, never guessed at; that is a design rule across every language here.

5. The file you were given is probably cp1251

Most Bulgarian subtitle files in circulation are Windows-1251, not UTF-8, and many exist only as frame-based MicroDVD .sub files. Read as UTF-8 they do not fail, they turn into replacement characters. The reader here scores candidate encodings on how plausible the decoded text looks, and reads the frame rate a .sub file declares in its first line rather than guessing one, because a wrong frame rate does not shift the timing, it stretches it.

How good is it, measured

Against a human Bulgarian translation of four episodes of a US sitcom, the pipeline scores 48.94 chrF++, the mean of three identical runs whose spread was 0.18. That number is a baseline for Bulgarian and nothing else: it is what any change to the prompt has to beat, and it is not comparable to the figure for another language. Cost to run was about $0.44 per episode in API calls, which is why an episode sells for $1.25.

Bulgarian is the language this was built for, and the page for it lists exactly what the profile checks. The first episode is free.