Just found this acoustic model (
zamia) looks very interesting, but I have some questions regarding it.
'bout;b'aʊt
'cause;kʌz
'course;k'ɔrs
'cuse;kj'uz
'em;ʌm
....
But the nonsilence phone it generated looks like (Sampa?):
'e: e:
'OI OI
tS
...
The problem is that the repo dose not provide an easy way to generate new phone for OOVs,
so I have to use kaldi's recipe to do the conversion while loading the repo's sequitur model and phonedict.,
but the kaldi's recipe (the create_dict.sh in librispeech, which i am using) seemed not handling
well this kind of symbols. What did i do wrong?
2> This repo seemed indifferently put all the corpus together (librispeech, ted, CV,...)and do the tdnn chain training from scratch (gmm)...
Just wondering is it the best way?
Thanks!