Search Strategies For Large-Vocabulary Continuous-Speech Recognition

This paper gives an overview of search strategies for large-vocabulary, continuous-speech recognition. We describe the adaptation of the one-pass beam search strategy to large vocabulary recognition. In the second part, we review the one-pass algorithm from the viewpoint of how the search space is organized and how duplication of partial hypotheses can be avoided. Thus we obtain natural variants of the conventional one-pass strategy: time conditioned copies vs. word conditioned copies and lattice vs. one-pass methods.