Research Of Voice Conversion Based On Frequency Warping Method

In modern society Voice Conversion is an important technology in speech signal processing, which modifies the personalized characteristics of a speech signal without changing its content. However, any modification applied to speech signals affect their perceptual quality. In particular, voice conversion on specific person involves prosodic and spectral transformations produces great quality degradation. A method represents a trade-off between the similarity and the quality of the resulting converted speech is of great importance. Among the current voice conversion methods, frequency warping offers a satisfying quality of converted result while the similarity is poor. The thesis researches on frequency warping methods in order to address the problems of current frequency warping methods. The main content of this thesis is as follows:(1) The thesis proposes a novel frequency warping method based on the bivariate mapping of formant.The problem of frequency warping method based on the formants frequency is the spectrum amplitude remains unmodified. In order to address the problems, the proposed method uses a frequency-warping function mapping both the formants frequency and formants spectrum amplitude of the source and target speakers. Compared to traditional formant mapping method, the method in the thesis achieved a more similar warped spectral envelope to the target spectral envelope, which improves the similarity score.(2)The thesis proposes a weighted frequency warping method based on GMM.The traditional frequency warping method results in inherent defects in the similarity of the converted speech. With reference to previous studies on the GMM and frequency warping methods, the thesis proposes a voice conversion approach using weighted frequency warping method based on GMM. The frequency warping function used in the proposed method is the bivariate mapping of formant frequency and corresponding spectral envelope amplitude of the source and target speaker. The mean advantage of this method is that it can improve the similarity of converted and target speaker’s speech as well as maintain a high quality of converted speech, i.e. it achieve a good balance between quality and similarity.(3)The thesis simulates related methods and analyses experiment results.This paper compares bivariate mapping of formant frequency warping voice conversion approach and regular mapping of frequency warping voice conversion approach, experiments show that the proposed method has a better similarity than the traditional method; The converted speech using weighted frequency warping method based on GMM, compare to classical GMM results, the method proved to obtain a better sound quality, at the mean time maintain the similarity.