Visual approach for automatic pitch period estimation

We present a visual and effective approach for pitch period estimation of recorded speech units. The approach is based on the recognition of pitch peaks and a comparison of the likelihood between the speech segments. The pitch peaks are composed of the 1st-order peaks and the 2nd-order peaks, which are determined by the original speech signal samples and the 1st-order peaks respectively. The average magnitude difference function (AMDF) on the 2nd-order peaks are calculated and normalized for comparing the likelihood between the speech segments. The pitch period estimation process is to select the appropriate threshold value of AMDF and the most probable 2nd-order peaks. This approach is computationally efficient, highly accurate and easily implementable. Besides, no need of training and reset of threshold before the approach is applied. The experimental results for the verification of our visual approach show that the average overall accuracy is over 99 percent