Swedish Subtitle Generator

Swedish is closely related to English, which makes it deceptively easy for a general model and quietly wrong in the details. The vowel system is large — nine vowels, each long or short — and several pairs sound close enough that a model guesses. The pitch accent that separates anden the duck from anden the spirit is inaudible to a model that was not trained on it. These are the errors that make a Swedish reader stop trusting the subtitles.

Särskrivning is the other giveaway. Swedish forms compounds by joining words, and writing them apart is such a recognisable error that it has its own name. English-trained tools do it constantly because English keeps its compounds separate. The transcript here writes compounds as single words, which is both correct and noticeably more readable on a phone.

The rest is regional and practical. Skåne vowels, the Gothenburg lilt and the more clipped Finland Swedish are all handled from the audio. English borrowings stay in English. And the export is UTF-8, so å, ä and ö survive every player rather than turning into question marks in the one that matters.