本論文結合語音合成及語者轉換的技術做相關的應用與設計，語音合成是真人聲音經合成引擎轉成機器音，語者轉換是以原來語者的聲音為基礎，轉換為另一語者的聲型發聲。要使這兩個技術能應用於生活及娛樂上，需要設計系統來供實作，本系統的設計為，輸入文字，經語音合成產生出來源語者的聲模，再加上文脈相依資料，經合成軟體，擷取出來源語者的頻譜特徵參數，再將目的語者語音擷取出頻譜特徵參數。兩者的頻譜特徵參數，經DTW比對，產生音框特徵向量匹配表，經LBG演算法形成高斯混合模型，用EM演算法，做高斯混合模型訓練，再經由GMM對應參數的方法，當輸入來源語者的頻譜參數，會轉出輸目的語者的頻譜參數，另外，激發出來源語者的音高特徵參數，再與目的語者的頻譜特徵參數，經合成濾波器形成目的語者的合成音。本論文提出語音合成及語者轉換之多項應用與設計。;The document combines speech synthesis and speaker conversion and these have relevant application and design. Speech synthesis is that the voice of real man is converted machine voice by synthesis engine. Speaker conversion is based on source speaker and it converts another voice of speaker. To let two techniques can be used in life and entertainment, it needs system to provide implement. The design of the system is that spectrum feather parameter of source speaker is extracted by synthetic software Data of text dependence produced by inputting words and voice model of source speaker input it. And the parameter of target speaker is extracted from voice of target speaker. Both of parameter generate the match table of feather vector of frame by DTW comparing, then GMM is formed by LBG algorithm. After that, using EM algorithm is order to train GMM. When finishing train, parameter correspondence method has transform function. When inputting source spectrum, target spectrum can be got. Besides, synthetic voice of target speaker is formed by speaker is formed by putting pitch feather parameter of source speaker excited and spectrum feather parameter of target speaker together through MLSA (Mel Log Spectrum Approximation). This document proposes many applications and designs of Speech Synthesis and Speaker Conversion.
