Optimal Segmentation for Analysis of Non-Stationary Voiced Speech