基于听觉模型特征的与文本无关说话人识别系统  

An Auditory Feature for Text-independent Speaker Recognition System

作　　者：卢小春[1,2] 尹俊勋[1] 王修信[2] 

机构地区：[1]华南理工大学电子与信息工程学院,广东广州510640 [2]广西师范大学计算机与信息工程学院,广西桂林541004

出　　处：《广西师范大学学报（自然科学版）》2010年第2期22-26,共5页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(60275005)

摘　　要：基于听觉生理和心理学的感知机理,提出一种适于说话人识别的特征参数提取方法。该方法采用Gammatone滤波器组代替常用的三角形滤波器组计算出语音信号各子带能量,根据等效矩形带宽模型,确定了各个滤波器的中心频率与带宽。同时从语音信号各子频带携带的说话人个性特征信息量角度分析对滤波器子带进行加权处理,最后结合倒谱均值减方法对特征进行处理,最终生成具有较高鲁棒性的特征。采用高斯混合模型对说话人进行建模,仿真实验的结果表明:提出的特征参数对含噪语音的与文本无关说话人识别具有较好的效果,在低信噪比情况下,具有比传统特征参数更好的噪声鲁棒性。The paper proposes a novel feature based on an auditory periphery model for robust speaker recognition. The sub-band energies of the extracted auditory features are calculated using a Gammatone filterband instead of commonly used triangle filter band. The center frequencies and bandwidths are then determined according to the equivalent rectangular bandwidth （ERB） model. Moreover,weighting the Gammatone filter bank by analyzing contribution of short-time spectrum in different frequency sub＇ bands,and using the CMS method to remove the varibility of channels are also investigated. Simulation results with Gaussian Mixture model indicate that the recognition accuracy is significantly improved by this auditory feature in the noisy environments for the text-indepentent speaker recognition,especially in low SNR environments.

关 键 词：听觉感知特性 Gammatone滤波器组 子频带加权 说话人识别 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]