Continuous Dynamic Patterns

The first chapter of this book gave an overview from a considerable altitude, with all the lack of detail that compels. The second chapter got stuck into practical issues of measurement of images as a sort of antidote, since in real life much hangs on the details and they need to be treated squarely. The third chapter was so theoretical as to alarm many a simple soul, but was necessary for the thinking man or woman in order to prepare them for a healthy degree of doubt about what follows, and some later discussion on neural models. The fourth chapter simply gave some convenient recipes for actually doing the job of recognising standard (static) patterns by statistical methods, and the last chapter explained how the standard neural nets work and how they accomplish pattern classification, when they do.

In the present chapter we start to deal with the problem of dynamic patterns, of things that change in time. Examples are speech, on-line character recognition, and such things as telling two missiles or two people apart, using knowledge of the way they move. I shall defer such issues as looking at video-images of moving objects later, until we have addressed the issue of how to work out what objects are in an image. In the present chapter I shall, in proper generality, discuss the issue of classifying trajectories in ${\fam11\tenbbb R}^n$ which arise from sampling some continuous process. In the next chapter I shall discuss issues arising from trajectories which occur in a discrete space of symbols known as an alphabet.

I shall start by discussing a simple and direct approach to Automatic Speech Recognition, the method being a staple of practical single word recognisers. This practical discussion will be followed by a short survey of the issues involved in continuous speech recognition. This problem is too hard to do much with at the present time, but the reasons it is so hard are interesting and worth thinking about.

The issue of noise in dynamic pattern recognition requires special treatment, it is something to which any measurement system is liable, and needs to be confronted. A great deal of engineering literature is concerned with this. I shall go briefly into the issues of stochastic processes or random time series and filters. I shall approach this from the engineering end rather than the statisticians end, although both have made contributions.

Since filtering theory and Automatic Speech Recognition (ASR) are both huge subjects in their own right, and quite impossible to cover in this book, the treatment will be confined to a survey of the basic ideas and pointers to the literature. The diligent reader will then, be in a position to write simple word recognition programs, will be in a position to understand and use samples of such programs on the accompanying disk, and will begin to understand the profound difficulties of the area.

Finally, I shall outline alternative approaches to Speech Recognition and discuss related problems.