一种使用声调映射码本的汉语声音转换方法

在使用高斯混合模型实现说话人语音频谱包络变换的同时，提出了一种汉语声调码本映射技术来进一步提高转换语音目标说话人特征倾向性的方法。从源语音和目标语音分别提取汉语单音节的基频曲线作为基频变换单元，作预处理和聚类后分别形成源、目标声调码本，根据时间对准原则建立了一个由源特征空间到目标特征空间的声调模式映射码本。声音转换实验评估了声调码本映射算法的性能。实验结果表明，该算法较好地反映出源说话人与目标说话人基频曲线之间的映射关系，改善了声音转换性能。

Voice Conversion with Tone Mapping Codebook of Mandarin Speech

With a mixture Gaussian mapping method for speech spectral envelope transformation, a Mandarin tone codebook mapping algorithm for the pitch conversion is proposed to improve the preference for target characteristics in the converted speech. The syllable pitch contours are first extracted from the source and target speech, and preprocessing and clustering operations are performed. A tone mapping codebook is obtained by associating the source codebook with the target one. Voice conversion experiments are performed to evaluate the tone codebook mapping algorithm. Results show that the method gives a good mapping between the source and target pitches, and improves the performance of voice conversion for Mandarin speech.