Voice Conversion Algorithm Based On The Acoustic Characteristics Of Personality Study

The voice conversion is to convert the personality characteristics of the source speaker into the personality characteristics of the target speaker while maintaining the original speaker's phonetic meaning and then synthesize it to make it similar to the target speaker's voice.In this paper,based on the study of the characteristics of personality acoustics,this paper improves and optimizes the shortcomings of traditional voice conversion methods in the transformation.The study contents include:1)For the Gaussian mixture is easy to occur the phenomenon of overly smooth when it transform the characteristics of voice as to degraded voice sound quality,we put forward a kind of the conversion model with the combination of Gaussian Mixture Model(Gaussian Mixture Model,GMM)and Generalized Regression neural Network(Generalized Regression Neuron Network,GRNN).GRNN is used to map the mean vectors of the GMM model parameter set and then replace them with the mean vector in the transformation rules formed by GMM model,at last we get a new mapping relationship.At the same time,the prosodic characteristic base frequency parameters are also transformed.Then the spectrum parameters and the base frequency parameters are combined to be transformed as the target voice.Finally,the experimental simulation and performance test show that this method can effectively improve the over smooth problem in the conversion.Compared with the traditional GMM model,this method has better speech quality and less distortion.2)It is easy to generate local convergence in the transformation of voicecharacteristics by the generalized regression neural network algorithm for particle swarm optimization.In this paper,a new Quantum Particle Swarm Optimization(QPSO)is proposed to optimize the speech conversion model of GRNN network.By changing the phase to cause the variation of the position vector,the algorithm is able to overcome the local convergence effectively.Therefore,the optimal smoothing factor parameters are obtained by using the quantum particle swarm to optimize the network,so as to establish the spectral mapping rule.Then,using the correlation between spectrum parameters and base frequency parameters,the prosodic feature base frequency is also transformed.Then the STRAIGHT model is used for synthesis.Finally,the experimental simulation and performance test show that compared with the traditional particle swarm optimization algorithm,the natural degree and similarity of the speech are improved,and the spectral distortion rate decreases.The proposed method has better speech conversion performance than RBF neural network,GRNN model and particle swarm optimization algorithm.