5th International Conference on Spoken Language Processing
As speech recognition systems are increasingly applied to real world problems, it is often desirable to use the same recognition engine for a variety of tasks of differing complexity. This paper explores the relationship between the complexity of the recognition task and the best strategies for pruning the recognition search space. We examine two types of task: 20000 word WSJ dictation, and phone book access using a 60 word grammar. For both tasks we compare two strategies for pruning the search space: absolute pruning, where the number of hypotheses is controlled by eliminating ones that have a score less than a fixed beamwidth below the best scoring hypothesis, and rank based pruning, where hypotheses are ranked by score and all hypotheses beneath a certain rank are eliminated. We present statistics characterizing the behaviour of the recognizer under different pruning strategies and show how the strategies affect error-rates.
Bibliographic reference. Balakrishnan, Sreeram V. (1998): "Effect of task complexity on search strategies for the motorola lexicus continuous speech recognition system", In ICSLP-1998, paper 0295.