A GRAMMAR COMPILER FOR CONNECTED SPEECH RECOGNITION

It is well known that syntactic constraints, when applied to speech recognition, greatly improve accuracy. However, until recently, constructing an efficient grammar specification for use by a connected word speech recognizer was performed by hand and has been a tedious, time-consuming task prone to error. For this reason, very large grammars have not appeared. We describe a compiler for constructing optimized syntactic digraphs from easily written grammar specifications. These are written in a language called grammar specification language (GSL). The compiler has a preprocessing (macroexpansion) phase, a parse phase, graph code generation and compilation phases, and three optimization phases. Digraphs can also be linked together by a graph linker to form larger diagraphs. Language complexity is analyzed in a statistics phase. Heretofore, computer generated digraphs were often filled with redundancies. Larger graphs were constructed and optimized by hand in order to achieve the required efficiency. We demonstrate that the optimization phase yields graphs with even greater efficiency than previously achieved by hand. We also discuss some preliminary speech recognition results of applying these techniques to intermediate and large graphs. With the introduction of these tools it is now possible to provide a speech recognition user with the ability to define new task grammars in the field. GSL has been used by several untutored users with good success. Experience with GSL indicates that it is a viable medium for quickly and accurately defining grammars for use in connected speech recognition systems.