特征空间本征音说话人自适应

提出了特征空间本征音说话人自适应算法,该方法首先借鉴RATZ 算法的思想,采用高斯混合模型对特征空间中的说话人信息进行建模;其次利用 子空间方法实现对特征补偿项的估计,减少估计参数的数量,在对特征空间精确建 模的同时,降低了算法对自适应数据量的需求.基于微软语料库的中文连续语 音识别实验表明,该算法在自适应数据量极少时仍能取得较好的性能,配合说话人自适 应训练能够进一步降低词错误率,其实时性优于本征音说话人自适应算法.

Feature Space Eigenvoice Speaker Adaptation

A speaker adaptation method at feature level named feature-space eigenvoice adaptation method is proposed. In this method, similar to RATZ, the information of speakers in the feature space is modeled by a Gaussian mixture model. Moreover, the number of parameters to be estimated is decreased by taking the dependency of these parameters into account. This method can use very little data to construct a more accurate feature space model. Experimental results of continuous speech recognition on Microsoft speech database show that this method can still achieve good performance even when the adaptation data is limited. And speaker adaptive training based on this method can further decrease the word error rate with a superior real-time performance to that of eigenvoice methods.