An Auditory-perceptual Pitch Detection Method

A new pitch detection method based on auditory-perceptual evidences has been proposed. As an auditory model, basilar membrane characteristics are approximated by using high Q pole and low Q zero pair filter with logarithmic frequency quantization. As the perceptual model distortion such as the missing fundamental problem and unvoiced segment interpolation problem are discussed. As a criterion function, harmonic structure evaluation function was used. An optimum pitch sequence is selected, using dynamic programming algorithm with continuity constraints, so as to maximize the sum of criterion value. Pitch detection experiments for Japanese utterances showed that the proposed method does not contain gross error and pitch data are interpolated smoothly in the unvoiced segment. These results show that the proposed auditory-perceptual model is effective and robust.