TheHidden Markov Model(HMM) has been widely and successfully used in speech recognition. However, it is difficult to design an HMM that operates in real time, as is required for automatic speech recognition or automatic target recognition. Instead of a conventional sequential computation environment, the use of a massively parallel computing environment for implementing an HMM should be explored. In this paper, a hierarchical neural model called theHidden Markov Learning Machine(HMLM) is proposed that successfully solves all three key problems concerning HMMs,learning,recalling, anddecoding. Because it provides a constant time response in the recalling phase, the HMLM provides the potential for real-time processing. The HMLM is constructed based upon the hierarchical neural model L3 [9],CTMAXNET, CTMINNET, labeling networks [8], and a special learning heuristics controller, REPLA [10]. An analysis of the time complexity of HMLM is also presented. The significance of this research lies in the progress from a stochastic model to a solid mechanism, similar to the progress fromPROLOGtoPROLOGmachines.

Keywords: neural networks, CTMAXNET, hidden Markov models, Viterbi algorithm, hierarchical neural model (L3), Baum-Welch reestimation algorithm

Received June 26, 1996; revised December 12, 1996.
Communicated by Hsin-Chia Fu.