摘要： 给出了一种基于声学分类的并行异方差PNN的说话人辨认系统。该模型的训练采 用最大似然准则，并发展了EM算法来调整参数。试验证明，新模型在保持辨认率的同时，也 减少了1/3的计算工作量。

关键词: 并行概率神经网络, 期望最大化, 说话人辨认

Abstract: In the paper, a speaker identification system that introduces acoust ic classification information into a heteroscedastic probabilistic neural networ k model is proposed. The training of the novel model utilizes maximum likelihood criterion and develops an effective EM algorithm to adjust model parameters. Th ese experimental results indicate the model maintains identification accuracy, w hile reduces the computational load about 1/3.

中图分类号: 

• TN912.3