在語音信號上的線性預測改良演算法

Abstract

我們從語音信號波形觀察，可以發現在語音出現的時間內，特別是在發母音的時候，一個取樣點的振幅，與鄰近取樣點的振幅有相關性。因此，某一個時間點的振幅便可以取其前若干點的振幅來做估計，此結果導出線性預測的概念。在語音信號處理中，線性預測是一項很重要的分析方法，實作上也常運用線性預測來取出語音的特徵參數。線性預測實際上是對口腔模型的另一種解釋，在線性預測中所得到的分析參數，可以用來轉換成描述語音在頻域上的特性。但是傳統的線性預測在分析上有以下的問題：在分析high-pitch的講者(通常為女性)共振峰時，傳統的線性預測方式會有不正常的極值出現在語音的包絡線上，造成較大的估計誤差；若預測的語音激發訊號有諧波的結構組成，在傳統的線性預測中其自相關函數會有膺頻效應的情況發生，一樣對語音共振峰的估計有不良的影響；此外，在處理受到雜訊干擾的語音信號時，傳統的線性預測表現也不盡理想。本論文研究整理改良語音信號線性預測的演算法，使其在估計語音共振峰的結果更為精確；另外，利用加權的方式，來降低雜訊對線性預測的影響；這些經過改良後的演算法經過推導證實，所得到的全極點濾波器皆為穩定。

We observe from the waveform of speech signal, and we can find out that in the period of voiced speech, the amplitude of one sample has a relationship with its neighbors. Therefore, we can estimate someone sample by taking previous other samples. And then, this result is the conception of linear prediction. In digital speech processing, linear prediction is a very important method of analysis, and we usually use linear prediction to get features in the speech signal in fact. By modeling the spectral envelope, linear prediction can capture the most essential acoustical cues of speech originating from two major parts of the human voice production mechanism, the glottal flow and the vocal tract. However, linear prediction analysis also suffers from some drawbacks, for examples, the biasing of the formant estimates by their neighboring harmonics which caused by aliasing that occurs in the autocorrelation domain and the phenomenon is most severe for high-pitch speaker in general. Additionally, it is well-known that the performance of LP deteriorates in the presence of noise. In this thesis, we try to improve conventional algorithm to solve these problems, in order to make spectral envelope estimation more accuracy and increase robustness against noise. By our verification, these improved algorithms all have the stability of the all-pole filters.