The power of adaptive (or neural network based) pre-processing has
been demonstrated in this project. However only a small fraction of
the available potential has been utilized, as evident from the
very high performances of the BASE_LINE system in training. There is
no doubt that the system has very good function approximation
properties. But the other properties required from a system used in
ASR, such as generalization ability, shift invariance, and temporal
pattern extraction ability, are yet to be improved. Typically the
improvement of one such property may lead to the degradation of
another. Therefore ways must be found by which all those properties
can be improved simultaneously, or at least one can be improved with
minimum effect to the others. In addition to system improvements, the
problem domain in which the system is intended to be used, has to be
widened. This project deals with only the isolated mode speech
recognition. Since the isolated recognition has only limited
applications, it is highly desirable that the system can be
modified to handle continuous speech. With those ideas in mind, some
suggestions for further work, has been given in the following.