论文部分内容阅读
在单芯片上实现的说话人确认系统是说话人识别应用的重要方向。该文面向片上应用,在使用DTW(dynamictime-warping)匹配方法的确认系统基础上提出一种改进的说话人确认算法,结合说话人确认的任务特点对DTW算法进行了改进:1)引入分层判决思想,2)在判决中结合单帧说话人区分能力估计,使系统的识别性能得到改进。新系统能够在对模板应用压缩处理后仍然保持良好的识别性能。测试表明新确认系统不做模板压缩时的等错误率为1.81%,经过1:8模板压缩处理后新系统的等错误率为2.35%。
The speaker recognition system implemented on a single chip is an important direction for speaker recognition applications. This paper aims at the on-chip application. Based on the DTW (dynamictime-warping) matching method, an improved speaker recognition algorithm is proposed, which improves the DTW algorithm by combining the speaker’s task features: 1) Judgment, 2) In the judgment, the capability of discriminating the ability of single speaker is improved to improve the recognition performance of the system. The new system can still maintain good recognition performance after applying compression to the template. The test shows that the new error rate is 1.81% when the new system is not compressed and 2.35% after the 1: 8 template is compressed.