Computer Science Department
School of Computer Science, Carnegie Mellon University


Incremental Network Generation in Template-Based Word Recognition


Kai-Fu Lee

December 1985

The network representation of word templates is presented. Using the network method, words are divided into segments, and same or different word templates can share segments. This not only reduces the storage required, but also enables the system to focus on acoustically dissimilar segments. Yet, by retaining multiple examples of the same word, it encapsulates variations of speech. A word recognition system has been designed and implemented to facilitate network training by providing 1 relatively reliable segmentation, 2 a segment-based warping algorithm that tolerates inexact segmentation, and 3 incremental network generation. Two network generation methods for isolated word recognition are presented, one of which achieved 99% accuracy for speaker-dependent alphabets, 92% for speaker-independent alphabets, and 99% for speaker-independent digits. Several other methods of reference generation were implemented using the same system, and the incremental network technique proved to be superior to all of them.

57 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by