机构地区: 华南理工大学电子与信息学院
出 处: 《微计算机信息》 2006年第08S期301-303,共3页
摘 要: 提出一种捆绑子空间分布隐马尔可夫模型的训练方法。该方法利用多变量相关系数将语音信号的特征向量进行子空间划分;利用k均值算法捆绑特征向量子空间的高斯分布,得到子空间高斯分布的原型,减少模型的参数。通过实验,用该方法训练的捆绑子空间隐马尔可夫模型,不仅提高了识别器的精确度和识别速度,而且节省了存储空间。 in this article, a new method to train Tied Subspace Distribution HMM (TSDHMM) is described. This approach is used to divide the acoustical observation vector space, tie subspace Gaussian distribution and get subspace Gaussian prototypes, which can reduce the number of parameters of a HMM. Evaluation on the task shows that a TSDHMM system achieves not only smaller memory requirement for acoustic models but also faster recognition without any loss of recognition accuracy.