An improved ANN method based on clustering optimization for voice conversion

Artificial neural network is a commonly used conversion model in voice conversion system, in which RBF is known for its concise convergence and fast learning. Based on optimizing the centers of RBF network, this article presents a method of using K-means algorithm to cluster and form centers and PSO algorithm to optimize the clustering number to improve the property of RBF, thus to enhance the transformation of speech parameters. Firstly, STRAIGHT model is used to extract linear prediction coefficients and pitch frequencies. Then the parameters are sent to RBF network, K-means and PSO algorithms are used to optimize the centers of RBF network until the fitness value is lowest. Experiment shows that, this method not only eliminates the trouble of finding the best clustering number one-by-one, but also effectively improves the performance of neural network, and the converted speeches are closer to the target one.