论文部分内容阅读
提出一种描述正常语音和变异语音之间关系的补偿因子。该补偿因子兼顾考虑了由于变异引起的特征分布中均值和方差的变化,并在k-均值初始化的参数基础上,采用期望最大化(EM)算法迭代估计变异补偿因子的值。通过估计出的补偿因子对变异语音特征进行补偿。对航空模拟飞行器中采集的应力变异下特定话者小词表孤立词的实验结果表明,利用所提出的方法可以将识别率提高32.3％。
A compensation factor describing the relationship between normal speech and variant speech is proposed. The compensation factor takes into account the variation of mean and variance in the feature distribution caused by the variation. Based on the parameter of the k-mean initialization, the value of the variation compensation factor is iteratively estimated by the expectation maximization (EM) algorithm. Mutation speech features are compensated by the estimated compensation factor. Experimental results on isolated words of a specific speaker in the simulated flight aircraft show that the proposed method can improve the recognition rate by 32.3%.