Learning frame dependencies in sequential data

In this paper, we propose a method to learn the dependencies existing among the frames for sequential data. We derive this method in the input/output hidden Markov model (IOHMM) framework. This method has the potential ability to increase storage capacity of hidden state to encode the past information as well as the capacity of an observation distribution. Based on this method, it is interesting to find that there exist a partial unification between IOHMM and the generalized fuzzy hidden Markov model (GFHMM). Then we implement a relatively more effective model called δ-IOHMM which can improve the performance without increasing any parameters. We apply the implemented model to speech recognition and compare the performance with the classical HMM and with the GFHMM. Some conclusions and promising empirical results are presented.