Piecewise linear stylization of pitch via wavelet analysis

In this paper, we propose a wavelet analysis based piecewise linear stylization of the pitch trajectory. We also address the often-faced difficulty in handling the tradeoff between mean squared error and the number of lines used for fitting, where a heuristic approach is typically used to make the stylization choice. We pose the piecewise linear stylization task as a minimization problem by defining a penalty function that is a linear combination of the stylization mean squared error and line number to seek an optimal tradeoff. The weights for the penalty function are selected in a semi-supervised way using a development set. We also provide an objective statistical measure based on such penalty function for evaluating the performance of the stylization problem. Results show that our algorithm provides 16.7% penalty reduction than the baseline system based on heuristics. Also, we found that the wavelet decomposition combination selection approach outperforms the low pass level selection approach by 15.6%.