一种改进高斯混合模型均值项的语音转换方法

摘要

语音转换技术主要应用于计算机语音合成、计算机语音翻译、语音编辑、广播及多媒体等方面。高斯混合模型（GMM）是目前语音转换的主流方法，但它的最大不足是会导致转换频谱的过平滑。其中GMM转换函数中的均值项和相关项共同导致了过平滑现象，并且均值项的影响更大。为此提出了结合码本映射法和GMM方法的修正均值法，实验表明，使用修正均值法能够有效抑制过平滑问题。改善转换性能。%Voice conversion has application in text to speech synthesis, voice editing, broadcasting and multimedia voice applications. GMM is a mostly used algorithm in the applications of voice conversion. However it causes overfitting in the converted voice spectrum which affects the transformed voice's quality. This paper analyzed this problem and found that it is caused by both of the mean value and covariance items in transformation function. To improve the performance of voice conversion, this paper proposed a new method combined codebook mapping method and GMM. Objective evaluations show that this method reduces the effect of overfitting, and improves the converted voice's quality.