As part of a MITRE internally funded R&D program in automatic speech
recognition, I have written a suite of programs for training ``Hidden
Markov Models'' using the ``Segmental K-Means Algorithm.'' This suite
includes shell scripts and programs written in C and Haskell.
The source code and some test data for the Haskell programs (and only
the Haskell programs) is now available by anonymous ftp from
nebula.cs.yale.edu
in the directory
pub/haskell/incoming
in the file
hmms-2.0.tar.gz
Untarring the file creates a directory tree with a root directory
called v2.0.
This software was described at the recent Dagstuhl Seminar on
Functional Programming in the Real World, May 16-20, 1994. In
addition, a copy of the viewgraphs and the draft of a report handed
out at the seminar are in the doc subdirectory.
To compile and run, you'll need hbc 0.999.4 or ghc 0.19. The
makefiles are initially set up for hbc. You can't use hbc 0.999.5
without modifying some of the programs and one of the data files
because of a bug in reading plain-text representations of data values
containing the association operator `:='. For ghc, Will Partain once
sent me a bug patch to ghc 0.19, an updated copy of the file
ByteOps_ap_o.o to replace the original in the library file
libHSrts_ap_o.a; you may need this too.
Without the C programs, this package does not comprise a complete HMM
design package. However, it provides examples of useful, numerically
intensive programs written in Haskell, and can serve as benchmarks for
compiler developers.
David M. Goblirsch ([EMAIL PROTECTED])
Lead Engineer, Signal Processing Center
The MITRE Corporation, 7525 Colshire Dr, McLean VA 22102-3481, Mail Stop W622
voice: (703) 883-5450 or (703) 883-5769 FAX: (703) 883-6708
