From: Mailing l. u. f. U. C. a. U. <kal...@li...> - 2011-08-02 00:53:37
|
Those tools have not been fully integrated-- they were created for use in some HLDA updates that are not yet finished and not yet in the public version. For decoding settings-- the only setting which you might want to tune automatically is the LM scale factor (there is no word insertion penalty, and the beam setting depends on how fast you need it to be). There is a script in egs/rm/s1/steps/decode_tri1_latgen.sh which is mainly intended for debugging lattice generation but which also computes WERs for a range of language model scores (it does this by rescoring the lattices with a range of WERs). You could easily modify this to automatically tune the LM scale. In my experience people don't use things like Rprop to tune LM scales as you only need a coarsely tuned value and anyway you already know the approximate range. As regarding these types of optimization methods-- we'll have to see if we have other problems that require them. People have different styles when it comes to using optimization methods. Dan 2011/8/1 Mailing list used for User Communication and Updates < kal...@li...> > Hello! > > I have a question about Kaldi's optimization algorithms (Optimizers, > Rprop). How do you use them? > Are there some tools, or plans for making tools, for automated calculation > of decoder settings (beam, lm scale factor, word insertion penalty, etc) ? > > > ------------------------------------------------------------------------------ > Got Input? Slashdot Needs You. > Take our quick survey online. Come on, we don't ask for help often. > Plus, you'll get a chance to win $100 to spend on ThinkGeek. > http://p.sf.net/sfu/slashdot-survey > > _______________________________________________ > Kaldi-users mailing list > Kal...@li... > https://lists.sourceforge.net/lists/listinfo/kaldi-users > > |