ABSTRACT
Traditionally, the development of speech pattern recognition systems has been attempted by using the pattern matching technology based on distance computation incorporating dynamic programming (DP). In this approach, an input speech pattern is represented as a sequence of acoustic feature vectors, compared with class models, each represented in the same manner as the input pattern, and then decoded to the model class closest to the input, in terms of DP-based distance. This simple scheme was practical and effective for implementing recognizers in the then limited computational environment, and indeed many noteworthy systems were developed for connected word recognition as well as basic isolated word recognition (e.g., see [22]).
