great result congrats to winning the micro bench

#1
by neurlang - opened

graphemes for this model are perfect.

in a future iteration, consider tweaking the phonemes further:

  • remove ɣ. this seems not a slovak sound. I don't know how often it occured in data.
    It cost you some mistake. It can be replaced by any of { k, ɡ, x, ɦ } on a case-by-case basis.
  • remove ŋ. this is nasal n, but in slovak probably there is heard no difference to ordinary n.

optional (up to you) - if it makes it better then keep it:

  • remove ʣ . Just let the model emit d followed by z.
  • remove ʤ . Just let the model emit d followed by ʒ.
  • remove ʦ . Just let the model emit t folowed by s.
  • remove ʧ . Just let the model emit t followed by ʃ.

I will try to make more sentence and higher quality sentence data for you. Insane how you used the gerulata/slovakbert, I didn't know it can be done. This has a great potential.

Sign up or log in to comment