摘要
基于听觉模型的特性,仿照MFCC参数提取过程,提出了一种基于Gammatone滤波器组的说话人语音特征提取方法。该方法用Gammatone滤波器组代替三角滤波器组求得倒谱系数,并且可以调整Gammatone滤波器组的通道数和带宽。将该方法所求得的特征在高斯混合模型识别系统中进行仿真实验,实验结果表明,该特征在一定情况下优于MFCC特征在系统的识别率,同时在Gammatone滤波器组通道数较高或滤波器带宽较小的情况下,系统具有较高的识别率。
In this paper, a novel feature based on an auditory model and Gammatone filter band is proposed for speaker recognition, which imitates the parameters extraction process of MFCC. The frequency cepstrum coefficient features are calculated using a Gammatone filter band instead of commonly used triangle filter band. Moreover, the dimension and the equivalent rectangular bandwidth of Gammatone filter band could be adjusted. Simulation results with Gaussian mixture model indicate that the recognition rate is significantly improved compared with MFCC in some condition, and the correct recognition rate is higher by more dimensions or smaller equivalent rectangular bandwidth.
出处
《微型机与应用》
2012年第1期37-39,共3页
Microcomputer & Its Applications
基金
国家科技计划基金资助项目(2008RR0003)
贵州省国际科技合作计划基金资助项目([2009]700109
[2009]700125)