摘要
在说话人识别平台上,针对碳键型电话产生的线性和非线性失真带来的影响,提出一个系统的补偿框架。通过研究电话信道对频谱包络的影响,从留数角度分析虚峰出现的原因,结合留数归一化方法和传统的倒谱均值减,对中间维的倒谱加以提升。实验结果表明,在训练环境和测试环境不匹配的情况下,该方法具有较好的顽健性。
Based on the speaker recognition system, the linear and nonlinear influences of carbon-button telephone channel are discussed. A compensation system is proposed to weaken those effects. By researching the frequency difference between telephone channel and traditional channel, the phantom formants are found, which can be reduced by normalization of its residues. Combined with CMS and ceptral lifter, the system is more robust under mismatch environment.
出处
《计算机工程》
CAS
CSCD
北大核心
2008年第13期84-86,共3页
Computer Engineering
基金
黑龙江省博士后基金资助项目
哈尔滨工程大学基金资助项目
关键词
电话信道补偿
留数归一化
自适应成分加权
倒谱均值减
telephone channel compensation
normalization of residues
adaptive component weighting
CMS