论文部分内容阅读
分析了语音转换研究中使用高斯混合模型映射算法时转换特征出现过平滑的问题,认为协方差矩阵估计不准确导致的转换特征细节信息的丢失是产生过平滑问题的主要原因,提出了使用码本映射和高斯混合模型共同转换声学特征细节的混合映射算法。此外提出了利用音素信息进行快速高斯混合模型训练的训练方法。客观评价表明使用音素信息的训练方法比常规方法性能指标平均提高了12．87％,而混合映射算法在使用音素信息的训练方法基础上比传统高斯混合模型转换算法性能指标提高了27．13％
The problem of over-smoothing of conversion features when using Gaussian mixture model mapping algorithm is analyzed. The loss of conversion feature detail information caused by inaccurate estimation of covariance matrix is the main reason for the over-smoothing problem. Hybrid Mapping Algorithm for Jointly Transforming Acoustic Feature Details by Mapping and Gaussian Mixture Models. In addition, a training method based on phoneme information for fast Gaussian mixture model training is proposed. Objective evaluation shows that the training method using phoneme information increases by 12.87% on average compared with the performance of the conventional method, while the hybrid mapping algorithm improves the performance index by 27.13% compared with the traditional Gaussian mixture model conversion algorithm based on the training method using phoneme information