In human speech, the timing function is important for determining its duration, stress and rhythm; however, little attention has been paid to these issues when building a speech synthesis system. In the human brain, the cerebellum plays a key role in the coordination, precision and timing of motor responses. We have developed a talking robot, which generates human-like vocal sounds using a simplified cerebellum-like neural network model as the timing function. The model was designed using the System Generator software in Matlab environment and the timing duration of trained speech was estimated using hardware co-simulated with a field programmable gate array board (FPGA). The timing information obtained from the co-simulation, together with the output motor vector, is sent to the talking robot controller in order to generate vowels of short, medium and long duration. Using this model for short-range timing of less than 1200 milliseconds, we verify that the short-range learning capability of the cerebellar-like neural network is applicable to the speaking robot for generating a human-like speech with prosodic features.
ASJC Scopus subject areas