Conclusions

From the results of the experiments described in the previous chapter, we can draw the following conclusions.

1. Combined optimization of HMMs, together with probability calculation and Pre-Processing Neural Nets, is feasible. This means that system converges to a local minimum of the error surface without diverging or oscillating.
2. As a basic rule, systems with larger number of free parameters have inferior generalization ability than those with lesser number of free parameters. That is, the difference in recognition rates for training and testing is larger, for the systems with larger number of free parameters.
3. Adaptive pre-processing can lead to better system performances both in training and testing, than non-adaptive or classical pre-processing does. However if the adaptive pre-processing introduces a large number of free parameters to the system, then the generalization ability of the system can be degraded.
4. Same effect of adaptive pre-processing cannot be achieved by increasing the neurons or the layers of the system with a fixed pre-processing part. This means that adaptive pre-processing is un-substitutable.
5. Existence of a recurrent loop can enhance the recognition rates. However it is most effective when placed with in the adaptive pre-processing stage. If it is placed at a point after a fixed pre-processing part , then there may be the cases where the system performances can become even poorer.
6. In the case of fixed pre-processing, Fourier transform gives better results than Hartley transform does. The reason may be the higher shift invariance associated with the Fourier transform.
7. Adaptation of the filter bank alone, can give comparable results with those obtained by adaptation of both filter bank and Hartley transform
8. Simultaneous adaptation of Hartley frequency or time indexes with Gaussian shaped filter bank does not lead to a system which converges. Even though a converging system can be obtained by replacing Gaussian filter bank with a rotated version of it, and by letting the Hartley time indexes adapt, the performances are not comparable with the best results obtained otherwise.
9. The system with an adaptive pre-processing part and a recurrent loop is the best, on the basis of recognition rates achieved both in training and testing. And the worst system is the one which has a Hartley transform based fixed pre-processing.
10. A hybrid HMM-ANN system with NN based pre-processing can give very good recognition performances and there is a big potential for further improvements

Home Page