EUROSPEECH 2001 Scandinavia
Decoding in a precompiled static network, compared with one in a dynamically managed network, is easier to implement and faster enough to yield a near real time response. However, when the recognition system handles a complex task, it has a problem of intensive memory usage. To overcome this weakness, we present a new decoding strategy that combines the advantages of static and dynamic network architectures. In this strategy, we first define a language model (LM) network that can represent an arbitrary back-off N-gram in a finite state network (FSN). The LM network enables constructing a precompiled static network and partitioning the whole network into subnetworks using LM histories. Then the recognition network can be dynamically created and destroyed on the subnetworkĄ6s basis. To make dynamic management of networks as simple as possible, we also devise a data structure for network representation that self-structures its nodes and arcs. The final decoder maintains subnetworks as needed, but does not need to maintain nodes and arcs. Experimental results show that this semi-dynamic management of networks dramatically reduces memory usage at the cost of less than 10% increase of recognition time.
Bibliographic reference. Ahn, Dong-Hoon / Chung, Minhwa (2001): "A one pass semi-dynamic network decoder based on language model network", In EUROSPEECH-2001, 1821-1824.