一种基于预测谱偏移的自适应高斯混合模型在语音转换中的应用

基于高斯混合模型（GMM）的语音帧谱包络转换算法容易导致转换后的语音谱包络过平滑、语音细节特征受损。通过对GMM中协方差的准确性与谱包络过平滑现象的研究，提出了一种基于预测谱偏移的自适应GMM建模方法。该方法采用平滑加权算法对目标谱的偏移进行建模，并根据语音帧信息自适应调节预测谱偏移项的比例系数，结合高斯混合模型共同实现对谱包络的转换。实验结果表明，该建模方法能够有效抑制转换后语音谱包络的失真现象，提高转换后语音的清晰度、自然度和可懂度。

An Adaptive Gaussian Mixed Model Based on Predictive Spectral Shift and Its Application in Voice Conversion

Voice conversion algorithm based on Gaussian mixture model (GMM) may result in the over-smoothing of spectral envelop and the damage of speech feature.By analyzing the relationship between covariance''s accuracy and over-smoothed phenomena,this paper proposes an adaptive GMM conversion algorithm based on spectral shift,which uses the weighted average algorithm to predict the converted spectral shift.Both the proposed spectral shift and the GMM are adopted to realize the appropriate converted spectral sequence.Moreover,the spectral shift proportion and GMM correlation are adaptively adjusted by using the spectral parameter.The experiment results show that the proposed algorithm can effectively alleviate the over-smoothing and improve the clearness naturalness and intelligibility of converted voice.