Akjava's picture
Update README.md
007b667 verified
metadata
license: mit
language:
  - ja

trying merged model ja004 and ja005 ratio almost (3:2)

Final Result: FAILD

voice is not merged,most of them ja004 voice,but sometime speak ja005

Train

  • The train and valid split is based on mora balance-V1 and word length.

Stage1

seems just echo

Stage2 (2024-09-14_04-08-31 )

Not listen all checkpoint but this model speak just separated voices

  • The train and valid split is based on word length and mora balance v1 (80:20).