ZZT演算法之應用於語音激發信號擷取

Abstract

在信號處理的研究領域中，語音處理占了相當重要的地位。人類語音的產生可模擬為聲門送出的激發信號，與口腔物理構造所形成之濾波器兩部份的迴旋積。對此過程之細節的研究與語音特徵參數的取得，可運用在語音合成、轉換等的諸多領域。本文中，我們將對zeros of z-transform(ZZT) 演算法及其於語音激發信號擷取的應用作討論。經過Z轉換之後，語音信號的zeros 得以在Z平面上展現其mixed-phase (在半徑1之圓內外皆有zero)的性質，並可憑此進行語音激發信號和口腔通道濾波器兩部份的分離。此外，根據ZZT圖形的研究，由群延遲函數(Group delay function) 所取得之相位資訊可以透過Chirp group delay方法得到大幅度的改善，藉以取得口腔通道濾波器的特徵峰值。所取得結果將與現有語音處理工具相比較，並測試主激發信號衰減(Attenuated Main Excitation ,AME) 方法對特徵峰值取得的改善。

Speech processing has been one of the major topics in the research field of signal processing. The process of speech production can be modeled as the convolution of glottal source excitation and vocal tract filter. The research in the details of speech production and the characteristic extraction can be applied in the fields such as speech synthesis and transformation. In this thesis, we discuss the zeros of the z-transform(ZZT) algorithm developed by Dr. Baris Bozkurt[1] and its application to the extraction of the excitation pulse in the source-tract model of human speech signals. After z-transform, the zeros of the speech signals can be represented on z the plane and the mixed-phase property is revealed, which would be used in source-tract separation. On the other hand, by the study of the ZZT plot, the phase information obtained from group delay spectrum could be well improved using the Chirp Group Delay. Moreover, we present the capabilities of formant tracking by ZZT, making a comparison between the performances of ZZT with other speech signal processing tools, and apply Attenuated Main Excitation(AME) for further improvement.