A Nonlinear Dynamic Modelling for Speech Recognition using Recurrence Plot - A Dynamic Bayesian Approach

The paper describes about a novel nonlinear feature extraction technique based upon recurrence plot(RP). This plot not only helps in visualizing the system dynamics but also can be quantified. The Recurrence Quantification Analysis (RQA) characterizes various aspects of a dynamic system and makes it a suitable technique for feature extraction. We have taken three prime quantification techniques namely Recurrence Rate, Entropy and Average Diagonal Length. The information about the system gets distributed in these quantities. Hence we need a model that is capable of taking into account the information from all the three RQA techniques. Dynamic Bayesian Networks (DBNs) can model these information very efficiently. For this purpose we have used Factorial Hidden Markov Model (FHMM) which is a special case of DBNs. The proposed method works well even in presence of noise when compared with the conventional technique.