基于码本映射和GMM的语音带宽扩展

发布时间：2018-03-22 08:35

本文选题：语音带宽扩展　切入点：高斯混合模型　出处：《北京理工大学学报》2017年09期 　论文类型：期刊论文

【摘要】：采用传统的高斯混合模型(Gaussian mixture model,GMM)进行语音带宽扩展时,会出现所估计的特征参数过平滑的问题,其主要原因是协方差估计不准确而导致扩展的高频特征细节信息的丢失,因此本文提出了码本映射(codebook mapping,CM)与高斯混合模型相结合的语音带宽扩展算法.提取高、低频特征参数,并训练高斯混合模型,基于高斯混合模型参数训练偏移矢量的码本;在扩展阶段,利用偏移矢量的码本将低频偏移矢量映射为高频偏移矢量,再将高频偏移矢量与高斯混合模型估计部分相加作为估计的高频特征参数.对利用该方法进行带宽扩展后的语音质量进行主观/客观评测.实验结果表明,相比传统的GMM语音带宽方法,CM-GMM合成的高频语音更接近原始高频语音,明显消除了高频过平滑现象.
[Abstract]:When the traditional Gao Si hybrid model is used to extend the speech bandwidth, the estimated feature parameters will be too smooth. The main reason is the loss of the extended high-frequency feature details due to the inaccuracy of covariance estimation. Therefore, a speech bandwidth expansion algorithm based on codebook mapping and Gao Si hybrid model is proposed in this paper, which extracts the high and low frequency characteristic parameters, trains Gao Si mixed model, and trains the offset vector based on Gao Si mixed model parameters. In the extension phase, the low-frequency offset vector is mapped to the high-frequency offset vector by using the codebook of the offset vector. Then the high frequency offset vector and Gao Si mixed model are added together as the estimated high frequency characteristic parameters. The subjective / objective evaluation of the speech quality after bandwidth expansion is carried out by using this method. The experimental results show that, Compared with the traditional GMM speech bandwidth method, the high frequency speech synthesized by CM-GMM is closer to the original high frequency speech, and the phenomenon of high frequency over-smoothing is eliminated obviously.
【作者单位】：北京理工大学信息与电子学院;
【基金】：国际合作研究项目
【分类号】：TN912.3

【相似文献】