Integrated context-dependent networks in very large vocabulary speech recognition

All the components used in the search stage of speech recognition systems – language model, pronunciation dictionary, context-dependent network, HMM model – can be represented by finite-state labeled networks. To construct real-time recognition systems, it is important to optimize these networks and to efficiently combine them. We present new methods that substantially improve these steps. We show that an efficient recognition network including context-dependent and HMM models can be built using weighted determinization of transducers [6]. We report experiments with a 463,331-word vocabulary North American Business News Task that show a substantial improvement of the recognition speed over our previous method [9]. Furthermore, the size of the integrated context-dependentnetworks constructed can be dramatically reduced using a factoring algorithm that we briefly describe. With our construction, the integrated NAB network contains only about 1:3 times as many arcs as the language model it is constructed from.