基于话者无关模型的说话人转换方法

Voice Conversion Based on Speaker Independent Model

提出一种基于话者无关模型的说话人转换方法.考虑到音素信息共同存在于所有说话人的语音中,假设存在一个可以用高斯混合模型来描述的话者无关空间,且可用分段线性变换来描述该空间到各说话人相关空间之间的映射关系.在一个多说话人的数据库上,用话者自适应训练算法来训练模型,并在转换阶段使用源目标说话人空间到话者无关空间的变换关系来构造源与目标之间的特征变换关系,快速、灵活的构造说话人转换系统.通过主观测听实验来验证该算法相对于传统的基于话者相关模型方法的优点.

A voice conversion method based on speaker independent （SI） model is proposed. Considering the phoneme information that commonly exists in every speaker＇s speech, an SI space described only by the phoneme information is assumed to exist. Gaussian mixture model （GMM） is adopted to model the distribution of the SI space, and the mapping relations from speaker dependent （SD） space to SI space are described by linear transformations. The SI model is trained by using speaker adaptive training （SAT） algorithm on a muhi-speaker database. In the conversion phase, the conversion function from source space to target space is quickly and flexibly built by joining the transformations from source space to SI space and SI space to target space. The advantage of the proposed method is proved by the results of some listening tests compared with two representative conventional methods.