论文部分内容阅读
利用HMM模型状态间的混淆度,提出了一种新的状态结构调整算法,使不同的状态可以共享相同的高斯混合函数,并在EM算法的框架下推导出对状态结构调整后的增加参数,即状态间权值的重估公式.并对非特定人进行大词汇量汉语连续语音识别实验,实验结果表明状态结构调整后的系统不仅优于基线系统,还获得了比传统的参数增加方法更高的识别率,由此证明了状态结构调整方法的有效性.
A new state structure adjustment algorithm is proposed based on the degree of confusion between the states of the HMM model so that different states can share the same Gaussian mixture function and derive an adjusted parameter for the state structure under the framework of the EM algorithm. That is, the formula for revaluation of weights between states.Experimental experiments on large vocabulary Chinese continuous speech recognition for non-specific people show that the state structure-adjusted system is not only better than the baseline system, but also obtains more advantages than the traditional method of parameter addition High recognition rate, which proves the validity of the state structure adjustment method.