Automatic Estimation of Fundamental Frequency Contour Parameters and Automatic Acquisition of Generative rules

点击翻译按钮获取中文摘要

(1) Smoothing and interpolation of measured F_0 contours of read speech of JapaneseMedian-smoothing for removing gross errors of pitch detection, linear interpolation of voiceless intervals, and recursive piecewise approximation of the resulting contour by third-order polynomials, were combined to obtain a mathematical approximation to the measured F_0 contour that is continuous and differentiable everywhere.(2) Automatic estimation of parameters using the derivative of the smoothed F_0 contourFirst-order approximations to the timings and amplitudes of the accent commands were obtained from the derivative of the smoothed F_0 contour, while those of the phrase commands were obtained from the residual. These first-order estimations were then refined by the method of Analysis-by-Synthesis to obtain optimum estimations.(3) Automatic acquisition of rules for prosody generationAutomatic acquisition of rules for prosody generation were investigated. From analysis results of a large amount of speech material obtained by the above-mentioned methods, it was found that a three-level quantization was perceptually acceptable both for the amplitudes of the accent commands and the magnitudes of the phrase commands for synthesis of read speech of Japanese.