Erik McDermott
Dynamic Programming for the Prototype-Based Minimum Error Classifier
Abstract:Currently, most speech recognition technology involves some form of Dynamic Programming
(DP) to cope with the non-linear compressions and expansions of speech event durations.
In choosing the particular version of DP now used in the Prototype-Based Minimum Error
Classifier (PBMEC) (McDermott and Katagiri, 1992, 1993, 1994), we aimed to use an approach
that was conventional, yet which also incorporated some of the latest advances in the
fast evolving field of search.
In this report we describe the use of one-pass DP (Sakoe 1978, Ney 1984), histogram based
pruning (Ney et al., 1994) and A* based N-best search (Soong & Huang, 1991, 1994), in the context of PBMEC. Our primary goal in this report is to explain these different methods
and their use in PBMEC. The motivation for choosing this particular configuration comes
from considerations of speed and search accuracy. In addition, we contrast two different
approaches to speeding up the search. The first, proposed by Soong & Huang (1994), consists
in using a simple forward phase with a reduced grammar, followed by a detailed backward A*
search using the full task grammar. The second is a simple time-synchronous pruning method
proposed by Ney et al. (1994). Our results suggest that the latter method is more effective
in reducing search time for the tasks we examined.
The recognizer used throughout this study is the Prototype Based Minimum Error Classifier.