Outline: Applications of Neural Nets Nettalk -learning Pronunciation of English Text Classifying Sonar Targets 16.1 Nettalk 16.1.1 Overview Phoneme String Text Speech Figure 16.1: a Text-to-speech System Using Nettalk

NETtalk is a classic example of a back-propagation trained multi-layer perceptron network applied to a practical application. NETtalk, created by Sejnowski and Rosen-berg 1], applies a multi-layer network to the text-to-speech problem. The goal is to develop a system which can convert English text into its underlying sequence of phonemes and stress markers. The string of phonemes and stress markers can then be used by a speech synthesizer to generate an audio realization of the text as seen in Figure 16.1. In using a network approach to the problem, it was hoped that NETtalk could learn a general mapping of spelling to pronunciation. Other current text-to-speech products such as DECtalk, utilize a dictionary lookup for common and irregular English words and apply a set of phonological rules to convert words which don't appear in the NETtalk DECtalk