https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    i think once you get to like sub-20 wavs you have to start relying on arpabet
  • m
    if so, 64 announcer on tacotron could be a thing if I did the same as I did with ultimate announcer
  • t
    Haven't been active enough to look into it but I will maybe by next week So is there another program I could use to maybe tackle this polyglot language machine design? Because I think I hit a dead end where I am.
  • h
    oh darn you're still having issues? hmm simplest thing that could possibly work: get any multispeaker tts model that takes phonemes as input, add both english and japanese phonemes into the "alphabet", yeet voices at it, pray that it generalizes across languages
  • c
    You want a multilanguage TTS? I mean, you train a voice and English and make it speak in Japanese?
  • t
    Yes, and vice versa
  • t
    I was working with a template from an English-Russian model
  • t
    Issue's a bit bizarre
  • t
    I'm in the synthesis step
  • h

    https://i.ibb.co/3NVGbbq/despair.png▾

  • t
    The process looks for a file called metadata.csv within a folder called russian_single (which I don't know how to change)
  • t
    I'm not sure if this actually is the issue, but I have a metadata file for every speaker
  • t
    But I think it's supposed to do the whole process in one go?
  • t
    Which means I would need to be using only one metadata file?
  • t
    Given that the creator used a different dataset collection for his voices, that might explain matters. I have maybe 550 separate metadata files myself.
  • t
    Anyway, that is jumping ahead, because anytime I try running the command I get a variety of errors
  • t
    This program I was using started with a g2p file. Any others that do that effectively?
  • t
    I have the g2p file
  • c
    I think I already sent you this, but I don't remember if you tested
  • c
  • c
    It has Japanese and English on it
  • t
    Oh I remember this I think it confused me when I tried the demo
  • t
    If I wanted to do English-Japanese, what would I do in the sample Colab?
  • t
    ...hm, maybe I shouldn't be looking at the demo
  • c
    You would need to train it
  • c
    By what I remember you specify the language in the transcription
  • t
    Sorry, but could you walk me through writing a notebook on this? The first few steps are clear until they ask me to apply the CSS10 repository
  • t
    For Japanese only (is English not a dataset because I don't see it listed as one)
  • c
    I will need to read more through this
  • t
    In the Read Me section
1...684685686...1068Latest