Context-dependent search in a context-independent network

This paper introduces a search algorithm for continuous speech recognition, working on a network that integrates both lexical and linguistic constraints. It differs from traditional Viterbi beam-search in that it does not assume that the network includes any information regarding context dependency of the acoustic models. Phonetic context dependency is instead taken into account by the search procedure itself, in a way that uniformly deals with within-word and cross-word contexts. In the paper the algorithm is described in detail, and results are given on two representative tasks: American English dictation and Italian broadcast news.