摘要
源-目标说话人声音转换是一种变换说话人声音特征的技术,它将源说话人的声音转换成目标说话人的声音.其中,声道参数的转换是获得高质量重建语音的关键,所以选择声道共振峰参数作为待转换的特征参数,利用线性预测求根法提取共振峰参数.为了克服分类线性转换算法(CLT)中分类不准带来的误差,引入了分类线性加权转换的策略,给出了一种基于径向基函数神经网络的分类线性加权转换算法(WCLT).在微软汉语普通话语音数据库上对转换语音分别作了客观和主观评估,验证了分类数目和训练集对两种转换算法的影响.实验结果表明,WCLT算法的转换效果优于CLT算法,一定程度上克服了高斯混合模型的转换算法(GMM)转换语音时,频谱过分光滑的现象,并在只有较少训练集数据时也能得到较好的转换效果.
Voice conversion is a method which transforms the source speech to a speech signal with the acoustic characteristics of the target speaker. The vocal-tract mapping algorithm is the key part, so formant parameters which are estimated by the root finding method based on LP analysis, are chosen for the transformation parameters. A classified linearly weighted transformation based on a radial basis function neural network was presented to reduce transformation error caused by inaccurate classification of classified linearly transformation. Objective evaluations and subjective evaluations were conducted in MSRA Mandarin speech database, and some experiments about the number of class and the training data were carried out. Experimental results prove that WCLT has a better performance than CLT, which can overcome the excessive smoothness of GMM, and the performance of WCLT has little bearing on training data.
关键词
声音转换
共振峰参数
径向基函数神经网络
分类线性转换
分类线性加权转换
voice conversion
formant parameters
radial basis function neural network
classified linearly transformation
classified linearly weighted transformation
