Post AtnmsvGLnAGvVShC08 by kazriko@alcatir.com
 (DIR) More posts by kazriko@alcatir.com
 (DIR) Post #Atms4iCUUKOP5AN59s by zaitcev@shitposter.world
       2025-05-05T15:00:23.782402Z
       
       0 likes, 2 repeats
       
       https://www.youtube.com/watch?v=ZMesv13cHdwI wanted to spread the goodness among my American friends and enabled translation on Youtube. Result is barely passable. Granted, it is astonishing that such a thing even works at all, and the generated output has a certain resemblance of the speech in another language. But the road forward is very long.
       
 (DIR) Post #Atmw4BKp56bEy5Gsmu by penguin42@mastodon.org.uk
       2025-05-05T15:19:46Z
       
       0 likes, 0 repeats
       
       @zaitcev My local Gemma3  instance gave me:> статической памяти процессор и памятьprocessor and static memory> соединяются множеством проводов 16are connected by a multitude of wires, 16> проводов соединяющие выводы а вместеwires connecting the outputs together
       
 (DIR) Post #Atmw4CHJZdO5tUdaxU by zaitcev@shitposter.world
       2025-05-05T15:45:04.808650Z
       
       0 likes, 1 repeats
       
       @penguin42 I'm sure it can be done better even with today's technology.
       
 (DIR) Post #AtnEeNF85C4BV611iS by penguin42@mastodon.org.uk
       2025-05-05T15:59:58Z
       
       0 likes, 0 repeats
       
       @zaitcev Is the Russian transcript text correct - or is thatmangled as well?
       
 (DIR) Post #AtnEeOACeziiM6ibg0 by zaitcev@shitposter.world
       2025-05-05T19:13:19.739462Z
       
       0 likes, 1 repeats
       
       @penguin42 The Russian subs are generally readable, but have issues.1. Lots and lots of small errors, like "ne" instead of "mne": Google cannot hear the guy correctly. And he does not mumble or anything, I can make out all the words.2. It ignores sentence structure and there's no punctuation at all.3. It has no semantic cross-checking, like Madonna thanking 40,000 Russian ventilators. For example: "processor laughs to read data from memory".4. Subs have significant gaps once in a while, as the whole thing just stops working. Maybe they would be generated in a batch mode for more popular videos.
       
 (DIR) Post #AtnmsvGLnAGvVShC08 by kazriko@alcatir.com
       2025-05-06T00:37:18.553397Z
       
       0 likes, 0 repeats
       
       @zaitcev I've used it before, you're right about it being barely passable. If you look at the native closed caption that it auto-generates you mostly know why though. If the voice transcription was more accurate it might deliver better results on the translation. The worst part is that they sometimes shadow-ban videos based on AI-misheard closed captioning results.
       
 (DIR) Post #AtnmswRjOKyBB31nNY by zaitcev@shitposter.world
       2025-05-06T01:36:55.076442Z
       
       0 likes, 1 repeats
       
       @kazriko It feels like it pattern-matches each word separately, without parsing phrases and sentences.