基于改进PNCC和i-vector的说话人识别鲁棒性

发布时间：2018-05-19 09:29

本文选题：幂率归一化倒谱系数 + 身份向量　；参考：《计算机工程与设计》2017年04期

【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-vector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10dB的噪声环境中具有更大优势。
[Abstract]:Aiming at the problem of insufficient robustness of traditional Mel frequency cepstrum coefficients (MFCCs) in speaker recognition systems, a method based on improved power-rate normalized cepstrum coefficients (PNCCs) feature algorithm and identity vector training model is proposed. Different from the traditional MFCC, the background noise is estimated by using long-time frames, and the robustness is further improved by multi-window spectral estimation, smoothing amplitude spectrum envelope and mean variance normalization. Based on the i-vector model, the speaker recognition experiment on the TIMIT speech corpus is carried out. The experimental results show that the proposed algorithm has the lowest equal error rate and the strongest robustness compared with other features under different noise and different signal-to-noise ratio (SNR). It has more advantages in noise environment where SNR is lower than 10dB.
【作者单位】：桂林电子科技大学信息与通信学院;
【基金】：广西区自然科学基金项目(2012GXNSFAA053221) 广西千亿元产业产学沿用合作基金项目(信科院0168)
【分类号】：TN912.34

【参考文献】