Genetic algorithm based RBF neural network for voice conversion

Voice conversion is a technique used to transform one speaker´s voice into another speaker´s voice. This paper describes a study on voice conversion using genetic algorithm (GA) to train the hidden layer of RBF neural network, which helps to improve the preference of the converted speech for the target speaker´s characteristics. Six mono-vowel phonemes in Mandarin speech are used for the conversion experiments which are performed on neural networks, respectively, by GA-based and K-means methods. Subjective and objective evaluations are conducted on the performances of converted speech. The conversion results show that in spite of not much improvement in perceptual distance, the RBF network by genetic algorithm instead of by K-means method has the ability of global optimization with an evident decrease in the spectral distance between the converted speech and the target speech.