• Captain Aggravated
    link
    fedilink
    English
    51 day ago

    I started to say this in my previous comment, but on things like Youtube shorts, I’ve noticed the baked in subtitles they always have tend to be hilariously inaccurate, even if the video is using a text-to-speech program to read aloud something written on Tumblr or Reddit, so they had the text in the first place… It does speech-to-text, then they run text-to-speech on that.

    LLMs are trained on written text, and I don’t think they would correctly innovate on misspelling. Someone else mentioned the “should of” mistake, which I can see an LLM doing, because it’s a common mistake humans have made. “cost” instead of “caused” isn’t commonly made by humans, so I don’t think an LLM would just come up with it. STT software has been pulling that shit for 30 years now though.

    • comfy
      link
      fedilink
      English
      21 day ago

      Absolutely. STT is still hit and miss on YouTube.