改进MCE训练算法在说话人识别中的应用

摘要

In practical problems, it adopts GMM - UBM as the background model when the training data is insufficient in speaker recognition system. Aiming at large amount of calculation in MCE training algorithm,it improved MCE. The improved MCE algorithm not only can reduce the amount of calculation, but also can get better recognition performance. The experimental results show that, under the different number of gaussian mixture and speakers, the improved MCE algorithm saves more training time than the traditional MCE algorithm, and as the growth of the number of gaussian mixture and speakers, the more time saving. In view of the MAP, MLLR, MAPMLLR and EigenVoice adaptation ways which used in speaker recognition system modeling, then using MCE algorithm and the improved MCE algorithm, the improved MCE algorithm has higher recognition rate than the traditional MCE algorithm.%针对实际问题中训练数据不足的特点,在对说话人建模时采用的是高斯混合模型—通用背景模型GMM-UBM,针对MCE训练算法中计算量大的显著问题,对其进行改进,改进的MCE算法不仅能使计算量减小,而且识别性能更佳。实验结果表明,在高斯混合数与说话人数不同的情况下,改进的MCE比传统MCE算法都要节省训练时间,且随着高斯混合数与说话人数的增长,节省的时间越多。针对采用MAP、MLLR、MAPMLLR、EigenVoice方法作自适应得到的说话人模型,然后应用MCE算法与改进的MCE算法,改进的MCE算法比传统MCE方法识别率更高。