Voice Conversion Using STRAIGHT Model And Deep Belief Network

Voice Changer semantic information is in the same premise, change voice personality characteristics make a person’s voice sounds like another person’s voice. At present, domestic and foreign scholars in the field of speech converting many studies, more classic methods: Based on the hidden Markov model(Hidden Markov Model, HMM), spectral folding(Frequency Warping, FW) codebook mapping(Codebook Mapping, CM) ANN(artificial Neural networks, ANN), and the use of Gaussian mixture model(Gaussian Mixture model, GMM) voice conversion. Most of which in the academic mainstream method is the use of domestic and foreign GMM model, and achieved good effect in voice conversion experiment.But the Gaussian mixture model for voice conversion process will be smooth over the phenomenon, but also a structural model of shallow Gaussian mixture model in nature. Shallow model has a feature that does not make mistakes in the model uses the premise of features directly determine the quality of the entire system performance, and dimension characterized generally only a few dozen dimension, can not adequately describe the characteristics between correlation and spatial distribution of the state, thus affecting the conversion performance.For these problems, this paper and the depth of belief network model STRAIGHT combination of methods to achieve speech conversion, the current depth of the belief network is mainly used in speech recognition. Firstly STRAIGHT model to analyze the speech synthesis, to avoid over-smoothing effect that occurs when using the voice conversion GMM model; secondly the use of networks to build the depth of faith-speech model, it has a more powerful than the "shallow" structural modeling and characterization capability to implement complex function approximation. Unlike shallow network, the characteristics of depth belief network is: 1) emphasize the depth of the model structure; importance 2) prominent feature of learning. Therefore, this method can be used for the speech signal features fully described minutiae voice reservations as to improve voice conversion effect.