Cost Reduction of Training Mapping Function Based on Multistep Voice Conversion

Several approaches based on a statistical method for voice conversion from one speaker to another have been developed. In a statistical spectral mapping method which is a typical one in these approaches, a mapping function which represents a correlation between different speakers is determined using spectral features. This technique has the problem that it is necessary to train the mapping function for each speaker pair. The training cost must become a serious issue in case that the number of speakers increases significantly. This paper describes a novel voice conversion method for reducing the training cost. This technique is easily implemented and can use conventional techniques directly. Experimental results demonstrate that the converted speech is almost maintaining the conventional quality despite the significant training cost reduction by the proposed method.