Design of a real-time hidden markov model state decoding system

Abstract

Hidden Markov Models are used in different kinds of sequence recognition problems. Specially, Hidden Markov Models are suited for speech/speaker recognition systems. Due to the complexity of the algorithms involved, general-purpose computing solutions are typically significantly slower than real time. For many applications, however, real-time is essential and thus a system based in specific purpose hardware becomes necessary. For the probability computation in pattern recognition systems using Hidden Markov Models, a state decoding system is necessary. The state decoding system must be able to decide, based on the input sequence, which is the most probable state sequence that produces the input sequence and therefore the reference pattern which best matches with the input pattern. In this work, the implementation of a real-time Hidden Markov Model state decoding system is described. The prototype was implemented for left-right Markov Models.