一种改进的新型说话人确认算法

Improved algorithm for speaker verification

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%。

Onchip speaker verification systems are a major application of speaker recognition technology. A baseline verification system using DTW (dynamic timewarping) matching is further developed with a 2level verification process and a frame discrimination method to improve the system performance. The improved system has good performance with template compression. Tests show that the equal error rates of the improved system are 1.81% without template compression and 2.35% with a 1∶8 template compression.