The DCD Library is a software collection for speech recognition
decoding and related functions. Based on the
Finite-State Machine
(FSM)
Library,
it provides higher-level operations needed specifically for
decoding. These include algorithms that:
make an optimized recognition
transducer (``network'') from a grammar, a lexicon,
and a context-dependency specification
search a recognition transducer to find the
best matching path or a set of paths (``lattice''),
given an input utterance and an acoustic model
weight lattices with just their grammatical and acoustic components
align reference and hypothesis automata using edit distances (``scoring'')
This library is not intended to provide a complete
speech recognition system:
there is no acoustic model, no acoustic feature generator (``front-end''),
and no acoustic, lexical or grammatical training provided by this
package; all are considered outside the scope of this library.
It does provide a general, extensible acoustic model interface,
it accepts general context-dependency, lexical and grammatical models
encoded as finite-state transducers, and it
reads pre-computed acoustic features in a variety of formats.
Our emphasis is on the generality, flexibility and performance of the
operations provided here, which are to be used as key algorithmic
components of an automatic speech recognition (ASR) system.
Several of the general operations should also find application outside of
speech recognition.
There is a program-level set of executables that is meant for use
in ``batch'' speech recognition training and testing. There is
a C++ library-level set of routines that is meant for easy integration
into an ASR system. The components are quite modular and independent,
so the user can select which functionality they require.
