A neural network application in thai text-speech conversion program
Hideaki Sugai, Stampe David · 1996
This dissertation demonstrates the application of neural networks in the field of linguistics, by employing them in a text-speech conversion task, specifically learning the value and classes of letters in the Thai writing system, and converting the letters into their phonemic representation. The Thai writing system presents an interesting problem in which rules based on traditional symbolic operation are not effectively deterministic. Based on theoretical and practical criteria, this study uses counterpropagation (CP) and the distributed method (DM) models. The performance of CP is always worse than the DM algorithm. The DM network, with 700 words trained and 60 training iterations, achieved 95.59% accuracy. The same DM network, with 800 words and 100 training iterations, achieved 90.57% accuracy. Then 100 unknown words are shown to the same network. The network produced 87.85% correct guess. In addition to converting consonants and vowels into phonemic forms, other tasks are applied: the leading vowel, the five tones and space components. Although these three tasks were not performed as well as the main consonants and vowels, a simple neural network can do a large part of the difficult tasks regarding intelligence without writing any specific symbolic rules. The neural networks developed for Thai reading in this study are considered practical and good tools to model 'mind-type' aspects of human behavior.