Hello! I actually work in localization (the creation of subtitles and dubbing) and first off I want to apologize because I think a lot of operators do lose sight of what these files are for and I am so sorry that it's ruined your watching experience. It's too easy to just start thinking of them as a deliverable.
But I also wanted to give insight into why these errors happen and encourage you to contact the streamers and/or production companies when you notice patterns in the file quality. Sometimes it's laziness, I will admit, but more often it's that we're scrambling to make the best file we can with limited funds and resources. Sometimes, too, these sorts of issues are evidence that a computer worked on the file, not a human - so pushing back on AI will help subtitles and translators stay employed and in turn give you a better viewing experience. It's well known in this industry that AI files are crap, but it's also something that studios don't put a lot of money into and will cut corners on with automation if you, the consumer, don't speak up.
For a bit of background, we work to a spec, which is the client's specifications for what can and cannot be included in the file, how to format different scenarios (phone calls, for instance), and how fast and how long events (each subtitle) can be.
speaks foreign language - I find this irksome too. Quite often this actually isn't included in the scripts (if we even get scripts, more on that later). Rarely do we get the translations, and I've never had them provide a transcript in the foreign language. If you see this frequently, I really urge you to write to the streaming company on which the show is playing.
shortening sentences - There is actually a reason for this, which also has to do with accessibility. We work to a certain reading speed and the files can only accommodate a certain length of the subtitle's event, so when a character speaks very quickly, sometimes paraphrasing is the only way of making it possible to read the event so quickly. Otherwise, you'd have to wind back and pause which would also mess up your viewing experience. A good operator will only drop extraneous phrases, names and such, and make sure the plot relevant sentence remains. Kids shows, especially, have strict limits on how long the event has to be on screen as, obviously, kids generally need a few seconds longer to read something than most adults.
changing what's been said for no reason - this can either relate to the above (sometimes the only way to shorten the sentence is to rephrase it) or reflect a change between a locked version of a script and the final version. It's not uncommon to have to fully transcribe without a provided script, and usually we are given what is called the locked video (a rough draft of sorts that may have different takes put in later or whole scenes removed). The final video follows much later and usually there is a very small window between that and the due date to get the files provided. A missing scene is very easy to catch, but the odd changed line unfortunately can slip through.
censoring swearing in captions - I can almost guarantee this is a spec issue and we aren't allowed to include them. A lot of specs explicitly instruct expletives to be removed. I really encourage you to contact the platform and encourage them to drop their censoring.
mumbles incoherently This isn't always in the script and, as mentioned above, we don't always get one. A good operator will take the time to go back over a sentence fifty times if that's what it takes to work out what the actor said. I've called coworkers over to have a listen to a line to get their advice, had heated debates over two different interpretations of what we think was said. Sometimes, however, the only way to get the file done is to describe exactly what you're hearing - incoherent mumbles - and move on.
Again, I'm so sorry. We're trying our best, and often for tiny budgets, tiny salaries, tiny turnarounds, and massive threats to our livelihoods from AI.
vaporwave
snowghoul
turnnoffyourmind
discoholicmusic
stormborns
tangentwithtanner
yousharknotpass
