摘要
在HMM的基础上,针对少量的训练样本情况,提供了一种新的训练算法—MCE(Minimum Classification Error)算法,并利用该算法建立了一个局部范围内不依赖于文本的说话人辨认系统,当每个说话人的样本训练量为10时,系统识别的正确率达到了97.14%。
In this paperwe present a new training algorithm-MCE(Minimum Classification Error)algorithm based on Hidden Markov Model(HMM)for little training data and use it to build a constrained text independent speaker identification system. When the number of training data is 10 for every speakeraccuracy rate of the system is 97.14%.
出处
《电声技术》
北大核心
2000年第11期3-6,共4页
Audio Engineering
关键词
说话人识别
语音识别
训练样本
MCE算法
speaker recognition Hidden Markov Model(HMM) minimum classification error