Abstract: Speaker transformation is that the speech of A is transformed to one of the B but the content of speech is not changeable.The pitch period is changeable during the pronunciation.The matching of feature parameters of two kinds of speech of speaker transformation applies moving window to extract the parameters of speech.Owing to the difference of period of speech signal，the matching error occurs.The proposed length-variable moving window algorithm is to select the different length moving window for speech segmentation，which caused each window to contain the same period of speech signal.Thus parameters difference caused by the difference of speech signal is eliminated.The experiments demonstrate the proposed algorithm enhance the performance of speech transformation.

Key words: speaker transformation, length-variable moving window, Gaussian Mixture Models（GMM）

摘要： 话者转换就是将A的语音转换为具有B发音特征的语音而保持内容不变。发音时基音周期是变化的，在语音转换的两话者特征参数匹配阶段，由于窗内语音信号周期不同，采用固定窗进行语音参数提取会造成了一定程度的匹配误差。提出的变滑动窗是按语音信号的基音周期变化来选择不同长度的滑动窗进行语音分割，这使得每个窗内的包含相同周期的语音信号，从而消除了由语音信号不同产生的参数差异。实验证明该方法提高了话者转换的效果。

关键词: 话者转换, 变滑动窗, 高斯混合模型（GMM）