Tarou Shirani
@taroushirani
sinsy自体はFLOSSなのにsinsy用の歌声モデルが沢山自作されて大盛況! とならなかったのはHMM自体が難解なためでは? (あとはhtsvoiceを作るのにマシンパワーが必要だったのもあるかもしれないが)
2020-11-07 14:42:16
Kanru Hua
@khuasw
This was exactly how I felt when first working on SPSS. HTK/HTS is a big hairy ball of 30-year-old code full of misleading and obsoleted features. HMM-based speech synthesis isn't even about HMM. It's about decision trees. So for the good of humanity use modern tools. twitter.com/taroushirani/s…
2020-11-07 15:44:51
Heiga Zen (全 炳河)
@heiga_zen
One needs read all codes of HTK first to be a master of HMM-based speech synthesis. IIRC I spent 2 months to understand the implementation including low-level memory management etc. twitter.com/khuasw/status/…
2020-11-07 15:54:01
Heiga Zen (全 炳河)
@heiga_zen
Note that I am not promoting it. To be honest, it was really tough to be a master of HTK/HTS :-P
2020-11-07 15:58:55
Kanru Hua
@khuasw
@heiga_zen Oh mine.. I can imagine all the pain. I didn’t even bother to use HTK from the beginning but spent two months reimplementing from scratch the core features of HInit, HRest, HERest, HVite, haha!
2020-11-07 17:23:22