Helmut Lucke
A study of Random Markov
Fields for Speech recognition
Abstract:In this report we argue that speech recognition is a two dimension pattern recognition process and that a speech recognizer could possibly benefit from using a two
dimensional stochastic process such as a Random Markov Field (MRF) instead of
an HMM. We address the problem of training such fields and present an algorithm
based on the EM-algorithm. This algorithm is experimentally evaluated for a unconstrained and a constrained MRF. We find that while the algorithm works well
in the unconstrained case there are some problems in the constrained version.