论文部分内容阅读
本文首先根据衡量话者特征参数有效性的F比值公式分析比较了LSP和CEP参数在描述话者特征方面的不同特性,确定LSP参数作为话者特征参数.接着提出了描述话者特征的层次模型.根据函数扩展网络的内插能力和隶属函数的模糊统计建立方法,用函数扩展网络实现模糊状态的隶属函数,由网络隶属函数构成话者特征层次模型的基本单元.以最大隶属度原则作为系统的辨认决策准则.在小词表(0-9十个数字)内,采用文本无关的方式进行42人的话者辨认实验,当测试语音由5个数字随机组合时系统正确辨认率为99.76%.
In this paper, we first analyze and compare the different characteristics of LSP and CEP parameters in describing speaker characteristics according to the F-ratio formula that measures the validity of speaker parameters and determine the LSP parameters as the speaker parameters. Then, a hierarchical model describing the speaker’s characteristics is proposed. According to the function of expanding the network interpolation ability and membership function of the fuzzy statistical establishment method, the function expansion network to achieve fuzzy state membership function, the network membership function constitutes the basic unit of speaker feature hierarchy model. The principle of maximum membership is taken as the systematic identification criterion. In the small vocabularies (0-9 digits), 42 people were tested in a text-independent manner. When the test speech was randomly composed of 5 digits, the correct recognition rate was 99.76%.