Sixth International Conference on Spoken Language Processing
This paper introduces a method for constructing word graphs in the extended decoding framework of m-gram language models (m > 2) and cross-word HMMs. The generation of word hypotheses contained in the graph relies on a word m-tuple boundary optimization extending the word-pair approximatlon. Two variants of graph generation are proposed: the first one fully encodes the cross-word and LM constraints used in the search into the graph structure which leads to compact sized graphs. The second method constructs word graphs with lower order constraints compared to those used in the search, resulting in larger graphs with lower graph error rates. Results are presented from systematic experiments carried out on the 5k WSJ and the 64k NAB tasks.
Bibliographic reference. Neukirchen, Christoph / Aubert, Xavier / Dolfing, Hans (2000): "Extending the generation of word graphs for a cross-word m-gram decoder", In ICSLP-2000, vol.4, 302-305.