Conversion Research, Based On The Excitation Source And Prosodic Features Of The Source - Target Speaker Sound

Voice conversion (VT) algorithms aim to modify the utterance of a source speaker to sound as if it was uttered by a target speaker, while preserving the original meaning.. VT is based on speaker verification and speech synthesis technology. Also, it enriches and extends these two speech research fields. Thus it is a valuable theoretical research topic and offers a number of useful applications. At present, many of VT's research works are focused on transformation of vocal tract spectrum parameters, however, the excitation source and prosody features, the main factors significantly affect the timbre of speech, are always neglected because of the their difficulties on extraction and modeling. In order to make the converted speeches have more target speaker's tendency, the thesis' works are mainly concentrated on transformation of excitation source and prosody features. Main contributions are summed as follows:1. For text-independent voice conversion, a probabilistic weighted transformation algorithm based on Gaussian Mixture Model (GMM) is utilized. The transformation algorithm performs a "soft" classification of source feature space by constructing a GMM that models the distribution of source feature space, which linearly combines a few rules derived from each subclass transformation. Thus the transformation accuracy is improved. The demonstrated empirically that the used method is more efficient and robust than VQ-based transformation method.2. Glottal source can affect the timbre of speech significantly. Normally, glottal flow derivative is used as glottal source. For the transformation of glottal source, a novel parameter extraction scheme for glottal flow derivative based on Legendre orthogonal decomposition is proposed. The algorithm employs the n...