摘要
介绍了一种基于连续M元高斯混合密度的隐马尔可夫模型(HMM)的非特定人孤立词语音识别仿真系统。通过研究模型状态数、训练时间以及特征参数选取对语音识别率的影响,得出HMM状态数取4,训练次数为20次,特征参数选取48维LPCC和MFCC的混合参数,可使语音识别系统对于汉语孤立词的识别率达到90%。
A kind of speaker-independent isolated word speech recognition system based on continuous M-ary gaussian mixture density of hidden markov model is introduced in this paper. Through analyzing the influence of training time, the number of model state and the selection of feature parameter on the recognition rate, it is concluded that the number of CHMM state is 4, the training times is 20, and a mixture of 48-dimensional LPCC and MFCC parameters as the characteris- tic parameters, the Chinese isolated word recognition rate can be up to 90% in the speech recognition system.
出处
《电声技术》
2013年第12期75-78,共4页
Audio Engineering
