基于改进PNCC和i-vector的说话人识别鲁棒性

摘要

针对传统的梅尔频率倒谱系数(MFCC)在说话人识别系统中鲁棒性不足的问题,提出一种基于改进幂率归一化倒谱系数(PNCC)特征算法和身份向量(i-vector)训练模型的方法.与传统的MFCC不同,PNCC利用长时帧估计背景噪声;在此基础上,通过多窗谱估计、平滑幅度谱包络和均值方差归一化(MVA)等技术进一步提升其鲁棒性.以i-vec-tor为基准模型,在TIMIT语音库上进行说话人识别实验,实验结果表明,在不同噪声、不同信噪比下,所提算法相比其它特征有最低的等错误率,鲁棒性最强,在信噪比低于10 dB的噪声环境中具有更大优势.%Focused on the issue that the robustness of traditional Mel frequency cepstral coefficients (MFCC) feature degrades drastically in speaker recognition system,a kind algorithm based on improved power normalized cepstral coefficients (PNCC) and bvector model was proposed.The difference between traditional MFCC and PNCC was that PNCC used long term frame to estimate background noise.On this basis,one way that using multiple windows spectral estimation,smoothing the amplitude spectral envelope and adopting MVA to enhance its robustness was proposed.The bvector was set as the baseline system for speaker recognition and test in TIMIT speech database.Experimental results show that for different noises and different signal noise ratios (SNR),the proposed method has the lowest equal error rate and the best robustness,and when SNR is lower than 10 dB,it has greater advantage compared to other algorithms.