Improved Automatic Extraction of Generation Process Model Commands and Its use for Generating Fundamental Frequency Contours for Training HMM-based Speech Synthesis.