• Leon@pawb.social
    link
    fedilink
    English
    arrow-up
    25
    ·
    3 months ago

    Can we start calling our own fuckups hallucinations and write off any sort of repercussions? “I’m sorry for not working for the past month, boss. I hallucinated that I was at work!”

  • KelvarCherry [They/Them]@lemmy.blahaj.zone
    link
    fedilink
    arrow-up
    23
    arrow-down
    1
    ·
    3 months ago

    In an example they uncovered, a speaker said, “He, the boy, was going to, I’m not sure exactly, take the umbrella.”

    But the transcription software added: “He took a big piece of a cross, a teeny, small piece … I’m sure he didn’t have a terror knife so he killed a number of people.”

    HUH??? What??? Huh???

  • friend_of_satan@lemmy.world
    link
    fedilink
    English
    arrow-up
    19
    arrow-down
    1
    ·
    edit-2
    3 months ago

    The astounding thing about this is that transcription should be deterministic. There’s no room for creativity here. Transcription is not a job that requires imagination. The closest you get to that is filling in stammers and stuff, but even that should be deterministic.

    • circuitfarmer@lemmy.world
      link
      fedilink
      arrow-up
      1
      ·
      3 months ago

      Despite mechanisms of linguistic change to avoid it, there are some instances of morphological ambiguity in natural language, meaning transcription is not always purely deterministic. It tends towards determinism the more narrow you make the transcription (e.g. for English, annotating stress patterns will avoid some ambiguity in places; transcribing to something like narrow IPA will avoid more), and context (=the semantic one) tends to handle the rest.

      That aside, trained humans can perform the task exceedingly well despite these pitfalls. The problem with the machine is it introduces its own extralinguistic issues at every level of analysis: just the necessary interaction of a phonetic model with an LLM adds a lot of slop (pun intended).