论文部分内容阅读
提出了一种共享空间旋转变换的声学建模方法。该方法结合状态空间旋转变换和方差部分共享的优点,克服了空间旋转变换方法由于每个输出都有一个变换矩阵而带来的计算量和存储量增加的缺点。在空间旋转变换方法得到比较精确的初始模型的基础上,通过共享的方差变换方法实现了不同状态的空间旋转矩阵的共享,解决了状态空间旋转变换后参数过多的缺点并提高了系统的识别率。试验结果表明,在汉语大词汇量连续语音识别系统中,同传统的对角方差建模技术相比,这种方法在计算量增加很小的情况下,系统字的误识率降低了18.8％。
A method of acoustic modeling of shared space rotation transformation is proposed. This method overcomes the shortcomings of the space-rotation transformation method that the computational complexity and storage capacity increase due to the transformation matrix of each output, combining the advantages of state-space rotation transformation and variance-sharing. Based on the initial model of the space rotation transform method, the space rotation matrixes of different states are shared by the shared variance transform method, the shortcoming of too many parameters after the rotation transformation of the state space is solved and the system identification is improved rate. The experimental results show that compared with the traditional diagonal variance modeling technique, this method reduces the misclassification rate of system words by 18.8% when the amount of computation increases is very small in Chinese large vocabulary continuous speech recognition system. .