[Kaldi-users] Kaldi updates

[Mailing l. u. f. U. C. a. U. <kal...@li...>]

Apologies to those who might have been expecting this alias to have more
traffic.

 I am just going to send out a brief update about a lot of Kaldi-related
stuff that has been happening.

This summer, in Brno University of Technology (in the Czech Republic), we
had a workshop for further developing Kaldi.  Attending were myself, Gilles
Boulianne, Arnab Ghoshal, Mirko Hannemann, Milos Janda, Stefan Kombrink,
Petr Motlicek, Yanmin Qian, Korbinian Reidhammer, Karel Vesely, and Thang Vu
(hope I haven't missed someone).

Things that we were working on included:
  Lattice generation (me, Mirko)
  Decoder work (Mirko, Gilles)
  Refactoring acoustic modeling code, and discriminative training (Arnab,
Petr, Yanmin, Korbinian)
  Neural-net LM rescoring (Stefan)
  Neural-net AMs (Karel)
  Tied-mixture systems (Korbinian)
  Switchboard recipes (Milos, with Martin Karafiat)

Not everything is finished yet.  Some of this stuff resides in the sandbox/
directory, in various branches.  The lattice generation is basically
finished, though, and there are some example scripts.

There is also somewhat more documentation available at kaldi.sf.net/, e.g.
the tutorial, and stuff on lattices.

As I probably mentioned before, and as documented on kaldi.sf.net/, we are
maintaining two versions that would be of interest to most people:
branches/1, which is the branch with the original functionality (relatively
stable), and trunk/, which is for new mainline development.  Soon we will
"tag" branches/1 as tags/1.0 or something like that, so there is a version
that won't change [this will probably be just before the ASRU conference,
where we will have a paper about the toolkit].

I have also been working on refactoring the example scripts, making them
cleaner.  Work on this is in the trunk/, in egs/rm/s3 (not 100% finished)...
eventually the plan is for all script directories to follow this pattern.  I
am hoping for the example scripts to include discriminative training and
hopefully tied mixtures within the next two or three months.

For the switchboard recipes, when we add them, there will be examples of how
to train LMs, and how to interpolate them (Gilles added this functionality
to IRSTLM in the form of a patch).

There is a recorded talk by me on WFSTs and lattice generation in Kaldi
available at
http://webcast.jhu.edu/mediasite/Viewer/?peid=8a2400f56d3841c79d96c46e3b2114391d.
 However, it is mostly an introduction to WFSTs from a mathematical point of
view and I don't claim it will be of interest to you.

Please let me or other developers know if you have any problems or would
like any features, or would like to contribute somehow.

Dan