采用STRAIGHT模型和深度信念网络的语音转换方法

Voice conversion using STRAIGHT model and deep belief networks

提出一种将STRAIGHT模型和深度信念网络DBN相结合实现语音转换的方式。首先,通过STRAIGHT模型提取出源说话人和目标说话人的语音频谱参数,用提取的频谱参数分别训练两个DBN得到语音高阶空间的个性特征信息;然后,用人工神经网络ANN将两个具有高阶特征的空间连接并进行特征转换;最后,用基于目标说话人数据训练出的DBN来对转换后的特征信息进行逆处理得到语音频谱参数,并用STRAIGHT模型合成具有目标说话人个性化特征的语音。实验结果表明,采用此种方式获得的语音转换效果要比传统的采用GMM实现语音转换更好,转换后的语音音质和相似度与目标语音更接近。

We propose a new voice conversion method which combines the STRAIGHT model with deep belief networks. Firstly, we utilize the STRAIGHT model to extract the speech spectrum parame- ters of the source speaker and target speaker which are then used to train the two DBN spectrum param- eters, and obtain the voice characteristic information of the higher order space. Secondly, we can con- nect and convert the two high order spaces using the artificial neural networks （ANNs）. Finally, we employ the DBN trained by the target speaker data to perform reverse processing on the converted fea- ture information, thus obtaining voice spectral parameters. Voice that has personalized features of the target speaker is synthesized by the STRSIGHT model. Experimental results show that compared with the traditional GMM based voice conversion method, the converted voice quality and voice similarity of the proposed method are closer to the target voice.