Improvements of procedures

In this project some limitations have been imposed in the procedures followed in testing and specially in training. Some such limitations, and their effects are listed below.

1. The training data base was limited to 100 sentences. This has an undesirable effect on generalization properties, and significance of estimated parameters. A larger data base, at least the TIMIT 462-sentence training set can be used instead.
2. Number of training epochs was limited to 100. Clearly we have risk that training has been stopped, before reaching a dominating local minimum. We can increase the number of epochs, sacrificing the training time. Precautions must be taken to avoid over training.
3. Whole training set was considered as a single batch. This leads to a less stochastic search path in the weight space and hence a more chance to be stuck in to a local minimum. This can be avoided by dividing the training set into several batches, and by operating on this sequence of batches

Home Page