基于Gammatone滤波器组的听觉特征提取

Auditory Feature Extraction Based on Gammatone Filter Bank

目前主流说话人特征参数在噪声环境中的鲁棒性较差。为此,提出一种可用于说话人识别的听觉倒谱特征系数。分析人耳听觉模型的工作机理,采用Gammatone滤波器组代替传统的三角滤波器组模拟人耳耳蜗的听觉模型,用指数压缩代替固定的对数压缩,模拟人耳听觉模型处理信号的非线性特性。在基于高斯混合模型分类器的识别算法下进行仿真实验,结果表明,该听觉特征具有比梅尔频率倒谱系数和线性预测倒谱系数更好的抗噪声能力。

Aiming at the problem that speaker＇s feature coefficients have poor robustness in noise environment, this paper proposes an auditory cepstral coefficient for speaker recognition. It analyzes the working mechanism of the human auditory model, simulates the auditory model of human ear cochlea by Garnmatone filter banks replaces the traditional triangular filter banks. Based on the nonlinear signal processing capability of human auditory model, exponential compression is used instead of the fixed logarithm compression. Simulation experiment is conducted based on Gaussian Mixed Model（GMM） recognition algorithm. Experimental results show that the auditory feature has better noise robusmess than Mel Frequency Cepstral Coefficient（MFCC） and Linear Prediction Cepstral Coefficient（LPCC）.