NeuroFuzzy approach to the development of a text-to-speech (TTS) synthesizer for deeply embedded applications

High quality text-to-speech (TTS) synthesis requires large amounts of computing resources (cpu and memory). To match deeply embedded applications we propose a novel approach based on soft computing methodolo-gies (fuzzy logic and neural networks). A feed-forward back propagation neural network has been trained for phonetic rules generation and a fuzzy logic engine has been tuned for prosodic control. A neural network has been used also to control the coarticulation process. A SOM neural network with fuzzy logic out-put layer has been defined for automatic utterance segmentation and labeling. Only basic phomemes units has been used for speech synthesis, demonstrating that a high-quality TTS synthesizer can be developed to target very deeply embedded systems.