基于类内类间距离的说话人特征优化

SPEAKER FEATURE OPTIMISATION BASED ON INTRA-CLASS AND INTER-CLASS DISTANCES

为了提高说话人识别的准确率,对常用说话人特征优化算法进行研究。针对加权系数法用升半正弦函数求加权系数没有考虑特征参量具体情况的缺点,提出一种基于类内、类间距离求加权系数的说话人特征优化算法。此算法对于具体的说话人识别选择更加适合的加权系数。对于各种说话人特征优化算法进行仿真实验,结果表明,改进算法对于说话人特征优化的效果更好,识别率可以达到非常满意的效果。

In order to raise the accuracy of speaker recognition, we studied the commonly used speaker feature optimisation algorithm. Due to the defects that the specific situation of characteristic parameters were not taken into account by the weight coefficient algorithm in obtaining the weight coefficient with raised half sine-function, we presented a speaker feature optimisation algorithm which is based on obtaining the weight coefficient by intra-class and inter-class distances. For particular speaker recognition, this algorithm could choose more suitable weight coefficient. We carried out the simulation experiment on various speaker features optimisation algorithms. Experimental results showed that the modified algorithm had better results in speaker features optimisation and the recognition rate could reach an effect with high satisfaction.