This package is a library and some associated scripts and executables for performing speech recognition using conditional random fields and segmental conditional random fields. It has been developed primarily with research in mind, as opposed to production speech recognition.

• OpenFST (dynamically linked)
• Quicknet3 (statically linked)
• cblas (dynamically linked)
• Kaldi installation (demo)
• TIMIT dataset (demo)
• autoconf 2.68 (development build)
• automake 1.15 (development build)
• gnulib/gnulib-tool (development build)

A source distribution packaged with autotools is available for download here. Download it to a location of your choice and deflate (e.g. with tar xvfz).

From the top-level directory (containing this README and the script configure), the basic installation is as simple as:

./configure
make
make install

Note, however, the following caveats:

Compilation depends on three external libraries for linking: OpenFST, Quicknet3, and cblas. If these are not installed in standard locations (e.g. /usr/lib/) on your target system, the configure script provides flags as a convenience for specifying their location, --with-libquicknet3-prefix=DIR and --with-libfst-prefix=DIR (see ./configure --help for details).

This package also comes with an optionally-installed demonstration of its Segmental CRF capabilities, for the TIMIT phone-recognition task. This demonstration depends on the existence of a functional installation of the Kaldi toolkit and the TIMIT dataset. If the location of the Kaldi TIMIT example recipe directory is provided at configure time via the --with-kaldi-timit-s5=DIR argument, the demonstration scripts will be installed. NOTE: this will silently overwrite the run.sh installed in that directory. Any important work in that file should be backed up before installation of this package. If you install the demo by specifying an argument to --with-kaldi-timit-s5, you must also specify the location of the TIMIT dataset, using --with-timit-data=DIR. DIR should be the top-level directory, containing test/, train/, doc/, and readme.doc. With the demo scripts installed, you can then run the script scrf-timit-demo.sh (which should have been installed to ${prefix}/bin). scrf-timit-demo.sh takes a single argument which is the name of a directory of your choosing, which will be used to store files generated over the course of the experiment.

This package has been built using autoconf version 2.68 (or higher) and automake version 1.15. You may be able to build it with older software, but it may take some coaxing. Some m4 macros were undefined when attempting to build using automake version 1.11, for example.

Dependencies: ASR-CRaFT statically compiles dependencies into a library and executables. It requires the QuickNet3 library (see above), CBLAS (for example, via ATLAS), and OpenFST (see above) to build. If these libraries and headers are installed in non-standard locations, locations can be provided to the configure script as described above for end-users.

The configure script depends on gnulib for a couple of m4 macros to facilitate finding library dependencies and setting appropriate rpath variables for the compiled executables. Thus, for development you will have to install gnulib-tool and invoke it as described in ${topsrcdir}/m4/gnulib-cache.m4. [Note that the gnulib source code contained in this package is distributed under the GPL in accordance with the Autoconf exception].

Ideally, you should be able to:

1. Clone the source
2. Run autoreconf in the top-level directory (you may need to run automake --add-missing first).
3. Run ./configure (see above about non-standard library and header locations; also, the .gitignore file is set up to allow a parallel build from a ./build/ directory. If you want to keep your source tree free of object files, you may want to create the build/ directory, and from build/, run ../configure [options]).
4. Run make.
5. Run make install.
6. Optionally, run make dist to package a new source distribution.