[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:02639] How to speed up the force alignment by HFst and OpenFst


Dear all,

I want to use HFst and OpenFst to align full-context label. But I find it cost large computation and memory.
It works like:

HFst -A -D -T 1 -C trn.cnf -I mlfFile -v 4.0 -w 1.0 -S train.scp -H cmp.mmf -N dur.mmf  -m hfst/cmp -r hfst/dur cmllist durlist

And then use fst2sym.pl, fstcompile, fstcompose, fstshortestpath, fstprint to get the lab. Just as the flow as HTS-demo 2.1.1

I tend to use the compact list for cmplist and durlist which are only for the current aligned file, but it's still cost too much. It seems build too large static network even reducing the phonelist. Is there any measures to speed up the alignment? For example, use '-c thresh' will do the favor?

Thanks 

--
Best regards!

Xi Wang (汪曦)

Beijing University of Posts and Telecommunications
Pattern Recognition & Intelligent Systems Lab

Follow-Ups
[hts-users:02659] Re: How to speed up the force alignment by HFst and OpenFst, Keiichiro Oura