节点文献
一种用于说话人辨认的概率神经网络的MCE训练算法
MCE-based PNN Training Algorithm for Speaker Identification
【Author】 Wang Chengru Wang Jinjia Li Jing (Department of Communication and Electronic Engineering, Yanshan University, Qinghuangdao 066004,China)(Department of Mathematics, Yanshan University,Qinghuangdao 066004,China)
【机构】 燕山大学通信与电子工程系; 燕山大学数学系;
【摘要】 提出了一种基于最小分类错误准则的概率神经网络的训练算法。实验结果表明,这种分类网络及其学习算法在30个说话人辨认应用中利用5秒清晰语音获得98.4%的辨认率,利用15秒电话语音获得85.1%的辨认率。
【Abstract】 This paper presents a training algorithm for probabilistic neural networks using the MCE criterion. Experimental results for speaker identification indicate that this classifier network and its training algorithm attain 98. 4% identification accuracy using 5 second clean speech utterances and 85. 1% accuracy using 15 second telephone speech utterances with a 10 speaker population.
【Key words】 Speaker identification Gaussian mixture model Probabilistic neural networks Minimum classification error;
- 【会议录名称】 第三次全国会员代表大会暨学术会议论文集
- 【会议名称】第三次全国会员代表大会暨学术会议
- 【会议时间】2002
- 【分类号】TP391.4
- 【主办单位】中国仪器仪表学会