论文部分内容阅读
本文首次使用了最大熵谱法估计的LPC反射系数矢量的长期平均作为说话人的语音特征矢量,利用所定义的特征矢量的平均自差异函数,平均互差异函数及平均互——自差异比函数分析了特征矢量用于非限定语音的说话人识别的有效性和说话人的可区分性;从模式识别分类的Bayes判决准则出发,导出了便于计算和程序实现的简化判决公式——欧几里德空间的加权和距离公式,并利用平均差异函数选择加权系数;提出了用序贯判别法对集外说话人的拒识方法;研制了相应的以微机为核心的实时响应的实验系统,响应速度为3秒。用此系统对20个说话人进行了非限定语音的说话人识别试验,误音率为10.67％,误拒率为5.67％,正确识别率95.41%。
This paper, for the first time, uses the long-term average of the LPC reflection coefficient vector estimated by the maximum entropy spectrum as the speaker’s speech feature vector. By using the mean autocorrelation function, the mean mutual difference function and the average mutual-self difference ratio function of the defined feature vector The effectiveness of feature vector for speaker recognition in unvoiced speech and speaker distinguishability are analyzed. Based on the Bayes decision criteria of pattern recognition classification, a simplified decision formula for easy calculation and program implementation is derived - Euclidean The weighted sum distance formula of German space, and the use of mean difference function to select the weighting coefficient; put forward the method of discriminating against the aggressors using sequential discriminant method; and develop the corresponding real-time response experimental system with microcomputer as the core, Speed of 3 seconds. Twenty speakers were tested by this system for speaker-free speech with unrestricted speech. The error rate was 10.67%, the false rejection rate was 5.67% and the correct recognition rate was 95.41%.