ABSTRACT
This chapter focuses on self-learning, connectionist approaches. It outlines the traditional, rule-based approach to text-phoneme conversion. Subsequently, a variety of automatic discovery techniques are treated, namely generate-and-test rule induction, rule induction by clustering, decision tree induction, Markov modelling, back-propagation networks in general, synthesis-by-analogy and syntactic neural networks. The standard approach to text-phoneme conversion in Text-To-Speech (TTS) systems uses a set of context-dependent translation (CDT) rules. The operation of rules has been described, and automatic-discovery techniques for inferring CDT rule sets reviewed. Synthesis-by-analogy has received renewed attention as a computational TTS model and as a single-route alternative to dual-route theory. This strategy of finding the best match to a lexically specified pronunciation stored in memory leads us to contend that memory-based reasoning (MBR) talk is using a form of synthesis-by-analogy. Lucassen & Mercer of the Hidden Markov model (HMM) techniques which were then starting to dominate speech recognition has inspired a small number of attempts to employ an HMM formalism for text-phoneme conversion.
