Next: Pruning
Up: Decoding
Previous: Decoding
A search procedure based on stack decoding [30,31] has been adopted. This search procedure may be regarded as a reordered time-synchronous Viterbi decoding and has the advantage that the language model is decoupled from the search procedure. Unlike time-synchronous Viterbi decoding, the Markov assumption is not integral to the search algorithm. Thus, this decoder architecture offers a flexible platform for single-pass decoding using arbitrary language models. The operation of the algorithm is described in some detail in [32,33]. Discussed below are some new approaches to pruning that have been developed to take advantage of hybrid system properties.