The development of a connectionist multiple-voice text-to-speech system

Experiments with a multiple voice text-to-speech system trained on the NETtalk and ATIMIT databases are described. Two connectionist models of the NETtalk architecture are used, one to do the letter-to-phoneme mapping and the other to map the phonemes onto LPC synthesizer control parameters. A comparison is made of the connectionist letter-to-phoneme system with a simple decision tree classifier, and it is concluded that the performance is similar, but the latter is more practical. It is noted that there is considerable difference between the letter-to-phoneme accuracy measured by exact reproduction of phonetic dictionary entries and the accuracy obtained by listening tests. Taking this into account, the system has performance comparable to other systems using hand-crafted rules