Voice Conversion Based On ANN

Voice conversion is a technology that converts one speaker's (source speaker) speech to be listened as if another speaker had spoken it, and it can keep voice content and tone invariable. This technology has important applications both at theoretic and practical practices. This thesis mainly focus on two key factors of voice conversion:pitch conversion and spectrum conversion. The main work and contributions are described as following:(1) It can be found that most of current conversion algorithms of pitch are simple linear conversion. But in fact, pitch conversion is nonlinear. Based on foregoing analysis, this paper proposes an algorithm of pitch conversion based on RBF artificial neural network. This method contains following steps:dividing pitch data into some equal segments, modeling those segments solely and finding morphing relations between them. Subjective evaluation and objective measurement indicates that the proposed method outperforms the traditional method both in synthesized quality and precision of mapping target characteristics.(2) Conventional conversion algorithms are used to convert parameters of segmental cues and parameters of super-segmental cues respectively. More and more studying reveals that relations lay in them. So this paper employs short-time combined spectrum as conversion parameters. Experimental results demonstrate that the performance of proposed method advance the traditional method in the same condition.(3) A conclusion can be easily drawn that the key factors influencing conversion performance of artificial neural network are means and power matrix of hidden layer. By reforming the key factors, conversion performance can be availably improved. Therefore, an algorithm that optimizing artificial neural network by quantum genetic algorithm is proposed in this paper. It is proved that the proposed method using quantum genetic algorithm has resulted in good perceived voices...