Experiments in Lexical Access for Speech Recognition

The Lexical access problem consists of determining the sequence of words that corresponds to a spoken utterance. In this paper, we present a Source-Channel model for addressing the lexical access problem, and describe experiments which investigate the impact of different choices regarding modeling and search parameters on performance. Our results show that the following yield significant improvements in lexical access performance: (1) using Dirichlet priors for estimating the back-off factor employed in a deleted-interpolation smoothing model, and (2) performing regional word adjustments during a post-processing stage. In addition, the use of short-lists of candidate words based on acoustic similarity significantly speeds up performance, with only a small drop in accuracy.