基于双因子高斯过程动态模型的声道谱转换方法

发布时间：2018-01-13 03:18

本文关键词：基于双因子高斯过程动态模型的声道谱转换方法　出处：《自动化学报》2014年06期 　论文类型：期刊论文

【摘要】：针对作者已经提出的双因子高斯过程隐变量模型(Two-factorGaussianprocesslatentvariablemodel,TF-GPLVM)用于语音转换时未考虑语音的动态特征,并且模型训练时需要估计的参数较多的问题,提出引入隐马尔科夫模型(Hidden Markov model,HMM)对语音动态特征进行建模,并利用HMM隐状态对各帧语音进行关于语义内容的概率软分类,建立了分离精度更高、运算负荷较小的双因子高斯过程动态模型(Two-factor Gaussian process dynamic model,TF-GPDM).基于此模型,设计了一种全新的基于说话人特征替换的语音声道谱转换方案.主、客观实验结果表明,无论是与传统的统计映射和频率弯折转换方法相比,还是与双因子高斯过程隐变量模型方法相比,本文方法都获得了语音质量和转换相似度的提升,以及两项性能的更佳平衡.
[Abstract]:Aiming at the two-factor Gao Si process hidden variable model proposed by the author, the Two-factor Gaussian process variable model is proposed. TF-GPLVM) does not take into account the dynamic features of speech when it is used in speech conversion, and the model training needs to estimate more parameters. A Hidden Markov model (HMMM) is proposed to model speech dynamic features. And the HMM hidden state is used to classify each frame of speech with probability soft classification about semantic content, and the separation accuracy is higher. Two-factor Gaussian process dynamic model, a two-factor Gaussian process dynamic model with low computational load. Based on this model, a novel speech channel spectrum conversion scheme based on speaker feature replacement is designed. Compared with the traditional statistical mapping and frequency bending transformation methods, or compared with the hidden variable model of two-factor Gaussian process, the proposed method can improve the speech quality and transform similarity. And a better balance between the two properties.
【作者单位】：解放军理工大学通信工程学院;解放军理工大学指挥信息系统学院;
【基金】：国家自然科学基金(61072042) 江苏省自然科学基金(BK2012510) 解放军理工大学预先研究基金(20110205,20110211)资助~~
【分类号】：TN912.3
【正文快照】： 语音转换(Voice conversion,VC)是在保持语义内容信息不变的前提下,通过改变语音中源说话人声音特征,使之具有目标说话人声音特征的处理技术[1].语音转换在个性化人机交互、信息安全、多媒体娱乐等领域有广泛的应用前景.例如,将语音转换和文语转换(Text-to-speech,TTS)技术

【共引文献】