A CORRECTING METHOD FOR PITCH EXTRACTION USING NEURAL NETWORKS

Pitch frequency is a basic characteristic of human voice, and pitch ex traction is one of the most important studies for speech recognition. This paper describes a simple but effective technique to obtain correc t pitch frequency from candidates (pitch candidates) extracted by the short-range autocorrelation function. The correction is performed by a neural network in consideration of the time continuation that is real ized by referring to pitch candidates at previous frames. Since the ne ural network is trained by the back-propagation algorithm with trainin g data, it adapts to any speaker and obtains good correction without s ensitive adjustment and tuning. The pitch extraction was performed for 3 male and 3 female announcers, and the proposed method improves the percentage of correct pitch from 58.65% to 89.19%.