Setting the target weights using acoustic distances

4 Setting the target weights using acoustic distances 

Various ways of setting the weights automatically have been proposed. Since the Hunt and Black algorithm is not probabihstic, we cannot directly use the most-standard probabilistic training technique of maximum likelihood. We can, however, investigate similar types of techniques that try to set the parameters so as to generate data that are closest to a set of training data. 

A maximum-likelihood approach would be to find the set of weights W = w, W2, .., Wp that generates utterances that are closest to the training data. To do this we require a distance metric that measures the distance between a synthesised utterance and the natural utterance. As we shall see in Section 16.7.2, defining such a distance is a difficult problem because it is in effect the same problem as defining the target function itself a good distance would be one at which utterances that are perceptually similar 

The solution to this then is to use a held-out test set, from which we leave out say 10% of the training data, and train on the remainder. When we now attempt to synthesize the 

The weakness of the left-out-data approach is that we are heavily reUant on a distance function, which we know corresponds only partially to hiunan perception. Various approaches have therefore been taken to create target functions that more directly mimic human perception. One way of improvii the distance metric is to use acoustic representations based on models of human perception for example Tsuzaki [456] uses an auditory modelling approach. However, as we shall see in Section 16.7.2, even a distance function that perfectly models human auditory perception is only part of the story the categorical boundaries in perception between one discrete feature value and another mean that no such measure on its own can sufficiently mimic human perception in the broader sense. 

---