README.txt 2011-02-22 Josef Novak

Transducersaurus is a module which builds component WFSTs for Automatic Speech Recognition Cascades (ASR). It contains classes suitable for building language model transducers from ARPA format LMs, lexicon transducers, context-dependency transducers and acoustic model to context-dependency mappers.

It provides both simple python prototypes and more robust as well as much faster C++ implementations of all the basic WFST algorithms needed to generate a standard LVCSR WFST cascade. It supports cascade generation for both the Juicer and TCubed WFST decoders. Sphinx support is on the way.

In order to create a sandbox and test the python tools, run the following, $ ./mk-sandbox.sh /path/to/my/sandbox $ cd /path/to/my/sandbox $ emacs -nw README.txt

Follow the instructions in the TESTING: section for a tutorial. See the prototypes directory for more details on the algorithm implementations.