First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

A Neural Network Approach to Multi-Language Text-To-Speech System

Yukiko Yamaguchi, Tatsuro Matsumoto

Information Processing Division, Fujitsu Laboratories Ltd., Kawasaki, Japan

This paper describes a neural network approach to phrase/clause boundary detection and letter-to-phoneme conversion that depend largely on the target language specification. The phrase/clause boundary detection network is provided with the parts of speech of three consecutive words and determines whether there is a phrase/clause boundary between the first and second words. The letter-to-phoneme conversion network converts input letters to distinctive features of the phonemes. The performance evaluation for English shows that the phrase/clause boundary detection network, trained with 500 sentences and tested with another 500 sentences, correctly detected the phrase/clause boundaries with 95% accuracy, and the letter-to-phoneme conversion network, trained with 1000 words and tested with another 1000 words, converted letters to phoneme correctly with 85% accuracy.

Full Paper

Bibliographic reference.  Yamaguchi, Yukiko / Matsumoto, Tatsuro (1990): "A neural network approach to multi-language text-to-speech system", In ICSLP-1990, 325-328.