• tal@lemmy.today
    link
    fedilink
    English
    arrow-up
    7
    arrow-down
    1
    ·
    edit-2
    11 months ago

    If anyone has an Nvidia card and a couple sound samples of Marge’s voice – you don’t need a ton of VRAM for this like you do Stable Diffusion, can do it on an 8GB card – you can fire up Tortoise TTS and render this in her voice.

    • Mechanite@lemmy.world
      link
      fedilink
      English
      arrow-up
      2
      ·
      11 months ago

      Is TTS voice replication much better self hosted than it was about half a year ago? Last time I tried it didn’t sound like the person I had samples of, instead I had to use elevenlabs to get close to sounding right

      • tal@lemmy.today
        link
        fedilink
        English
        arrow-up
        2
        ·
        edit-2
        11 months ago

        I mean, that’s a subjective question. I think it’s decent. Here’s some samples:

        https://nonint.com/static/tortoise_v2_examples.html

        Last time I was running it, Tortoise TTS doesn’t have a way to directly annotate voice with intonation or emotional stuff. The best you can do is pulling tricks like using a feature that lets you add some words to a sentence that aren’t actually spoken to affect the emotional impact of the words that are (e.g. sad words to make the spoken words be spoken in a sad voice).

        Imagine the difference between someone saying gloatingly “none of you will survive” and someone saying it in an agonized voice.

        I do wonder a bit whether it’d be possible to train it on a corpus that’s been automatically annotated with output from software that does sentiment analysis on text, and then generate keywords that one could use to alter the sound of sentences. I don’t think that this is so much a fundamental limitation of the software as it is limitations in the training set.