https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    yeah, i'm glad it worked!
  • u
    i'm going to continue training it and when it's done we'll have a french base model
  • z
    how much data? and, what did you use for transcriptions? IPA?
  • u
    i couldn't count how many minutes of data because the large amount of wavs causes a stack overflow in audacity but it's 6075 wavs
  • u
    i used ipa with word separation and punctuation for the transcription the cleaners were set to just return the text
  • z
    nice
  • z
    I wonder if we should move completely to IPA
  • z
    might let us train cross language base + multispeaker models
  • u
    it might work you could use the espeak phonemizer and yeah it would work regardless of the language
  • i
    what if bob was planning to do that with IPA already 👉 👈
  • i
    me and cris only did some small attempts so far and i see potential
  • c
    il semble avoir du mal à prononcer le mot poète
  • c
    (j'entends pote)
  • c
    That's what I've used to train the base model
  • c
    There's only one problem, after some commits on the Uberduck Pipeline, it isn't training on IPA properly anymore
  • u
    oui, je dois faire plus de training c'est encore pas très précis
  • c
    mmmmm combien d'epochs
  • u
    i used the normal tacotron training notebook and just modified symbols.py and cleaners.py to accommodate
  • u
    13, environ? bien sur pas assez
  • m
    Great idea, but I think hecko has a good idea to go along with it: add a small layer in the model to indicate the language (using an ID of some sort)
  • m
    Similar to speaker embedding
  • h
    coqui apparently uses a mere 4 dimensions for the language embedding by default
  • h
    granted they only have 3 languages in their multilingual model so
  • h
    also google used some sort of adversarial loss to disentangle speaker identity from language, not sure about the specifics
  • h
    it's 17 and 1:54 mins of data
  • t
    Can you extend the dataset to 50 samples?
  • t
    I think it is too low.
  • h
    i think it's decent enough
  • h
    especially since they're long samples
  • h
    I can try to space some stuff out a bit too I need to do it anyway to get the pacing a bit more accurate
1...908909910...1068Latest