变帧长和变帧率在说话人确认中的应用

Application of variable frame length and frame rate in speaker verification system

从变帧长、变帧率角度考虑提出一种新的提取MFCC的方法。该方法先将帧长和帧率都限制为基音周期的整数倍,即基音同步算法;然后基于变帧率算法的原理在语音特征变化缓慢的地方去除一些帧来降低帧率。在NIST 99说话人评测上进行的说话人确认实验表明,该方法不但提升了系统性能,而且降低了帧率,节省了特征文件的存储空间。

A new method for extracting Mel-Frequency Ceptral Coefficients （MFCC） was proposed from the perspective of variable frame length and frame rate. The proposed method restricted the frame length and frame shift to multiples of pitch period, called pitch synchronous algorithm; then removed some frames where the acoustic feature changed slowly to decrease the frame rate according to the principle of variable frame rate algorithm. With speaker verification experiments on NIST 99 speaker recognition evaluation, the new approach not only improves the system performance but also decreases frame rate, which means saving the storage space of feature files.