Learnable Phonetic Representations in a Connectionist TTS System — II: Phonetics to Speech

Abstract

In an earlier chapter, we described the overall structure of the SOMtalk text-to-speech system and detailed results suggesting that non-symbolic (‘phonetic’) representations — based on trajectories through a ‘phonetic’ space derived from a self-organising map — may play a useful part in deriving pronunciations from text. A similar strategy suggests itself for the subsequent stage in which synthetic speech is produced from the ‘phonetic’ representation. This makes it possible to bypass a symbolic ‘phonemic’ stage in the overall, trained system. In this case, only a small database has been used for learning because of the high computational cost of training on spectral data, but some encouraging preliminary results have been obtained.