Training of lexical models based on DTW-based parameter reestimation algorithm
Yoshiharu Abe, K. Nakajima · 2003
A systematic method for development of word models for large vocabulary word recognition is described. The word models are comprised of successive clusters of states whose durations are governed by continuous densities. They are generated by a lexical rule and trained by an iterative algorithm based on maximum likelihood estimation and DTW temporal alignment. Several lexical rules, along with conventional DTW matching method, are experimentally evaluated by both close vocabulary and open vocabulary tests in similar word environments. The results show the effectiveness of stochastic modeling and systematic generation of word models.>