基于循环最大期望修正算法的与文本无关话者识别研究

话者识别中目标模型的最大期望算法存在着出现奇异阵的重大缺陷，而最大似然估计虽然不会出现奇异阵，但识别率比较低。提出了一种循环最大期望修正算法，采用最大似然估计所得模型为初始模型，然后用最大期望算法中每步的模型，通过α值控制修正比例对其进行修正。实验结果表明，该修正算法较好地克服了奇异阵的出现，同时提高了识别率。

Text-independent speaker recognition based on recurring expectation maximum adjustment algorithm

For the expectation fantastic array, the targeted model in speaker recognition maximum algorithm, and although the there is fatal default of fantastic array about maximum likelihood estimate can＇t appear but there is lower rate of recognition. A recurring expectation maximum adjustment algorithm is proposed to utilize the maximum likelihood estimate to gain the initial models. These initial models are modified according to controlling adjustment rate with every model in the expectation maximum algorithm. Then more optimal models can be obtained. The results of the experiments show that the adjustment algorithm can conquer fantastic array well, and improve rate of recognition.