Learnable Phonetic Representations in a Connectionist TTS System — I: Text to Phonetics

Abstract

Results from connectionist experiments in text-to-speech conversion suggest that non-symbolic intermediate (‘phonetic’) representations may have a useful part to play in the design of a synthesis system. A similar strategy suggests itself in the subsequent stage when speech is produced from the intermediate representation, which makes it possible to bypass a symbolic, phonemic stage in the overall system, once trained. (This second stage is dealt with in a later chapter.) Error can still be calculated in terms of phonemes correct, but this is not necessarily a good measure of the naturalness and acceptability of the output speech. In contrast to other trainable text-to-speech systems, emphasis is laid here on the fundamental importance of phonetic and phonological sources of variability, and their separation from the underlying physical and temporal events. As far as possible, this phonetic/phonological capability should be built into the system prior to training on the main task at hand, as this corresponds more closely to the way these skills are acquired in human beings.