An improved method for voice conversion based on Gaussian mixture model

摘要

Voice conversion is a technology by modifying the source personality to mock the voice of target. This technology has a wide prospect and potential technical value both in technical field and entertainment, such as text to speech (TTS) and toys. This paper develops an improved voice conversion method. For the voiced frames, the conversion is implemented by Gaussian mixture model (GMM) based on speech transformation and representation using adaptive interpolation of weighted spectral contour (STRAIGHT) algorithm. For the unvoiced frames, the envelope is stretched or compressed according to the ratio of the vocal tract length (VTL) of the source and the target. The subjective experiment shows that the proposed method indeed improve the quality of the converted voice with the introduction of VTL.