Captions follow the language of the whole video, not its first 30 seconds
The spoken language used to be read from the first 30 seconds. A stream in Portuguese that opened with an English intro came out captioned in English. It is now read at seven points across the video.
Until November 19, Katto read the spoken language from the first 30 seconds of a video. A stream that opens with a countdown, a song or a few words in English was then transcribed as English from start to finish, even when the next three hours were in Portuguese.
On videos longer than 5 minutes, the language is now read at seven points spread across the video, and the majority wins. Of the 8 videos we had with the wrong language, 6 now come out in the right one. On 10 videos that were already right, the transcript is unchanged for 99 to 100 percent of the words.
What it does not solve: a stream that really alternates between two languages can still land on the one you did not want. In that case, choose the spoken language yourself in the settings before sending the video.
2 videos a month, no card. What is still being built is on the roadmap.