Speech Synthesis Using Neural Networks Trained by an Evolutionary Algorithm

This paper presents some results of our research regarding the speech processing systems based on Neural Networks (NN). The technology we are developing uses Evolutionary Algorithms for NN supervised training. Our current work is focused on Speech Synthesis and we are using a Genetic Algorithm to train and optimize the structure of a hyper-sphere type Neural Network classifier. These Neural Networks are employed at different stages of the Speech Synthesis process: recognition of syllables and stress in the high level synthesizer, text to phonemes translation, and control parameters generation for the low level synthesizer. Our research is included in a pilot project for the development of a bilingual (English and Romanian) engine for text to speech / automatic speech recognition and other applications like spoken e-mail, help assistant, etc.