带背景噪声的声纹识别系统的研究

发布时间：2018-04-17 23:36

本文选题：声纹识别 + 小波包变换　；参考：《哈尔滨理工大学》2014年硕士论文

【摘要】：本文所研究的声纹识别系统主要分为端点检测，特征提取和识别模型三个部分。端点检测部分主要研究了基于线性预测倒谱距离和短时过零率的双门限法，实验证明新的双门限法能够解决传统双门限法不能检测能量低的语音段的问题。特征提取部分，采用了美尔倒谱系数与差分美尔频率倒谱系数相结合的特征参数，更好的体现了说话人的个性特征。然后对高斯混合模型进行了研究，，提出了分裂法与K均值聚类法相结合的模型参数初始化方法，并用高斯混合模型对两种端点检测算法、特征提取算法和训练方法进行了仿真实验，在纯净的语音环境下，系统具有良好的识别效果。声纹识别研究的难点之一，即是在背景噪声下的识别系统的研究。虽然在纯净语音环境下的识别系统性能很好，但是在噪声环境下，识别率明显降低。本文运用小波变换和小波包变换对噪声进行了去噪处理实验，小波包去噪效果明显优于小波变换，然后在小波包常用阈值和折衷阈值的基础上提出了改进的阈值去噪方法，通过对语音信号的对比仿真实验，和对整个系统的实验数据表明，本文提出的基于小波包改进阈值算法很好地去除了噪声，去噪之后的识别系统取得了较高的识别率。最后将算法应用在实际复杂的噪声处理中，算法仍然有效地去除噪声。
[Abstract]:The voiceprint recognition system is mainly divided into three parts: endpoint detection, feature extraction and recognition model.In the end detection part, the double threshold method based on linear predictive cepstrum distance and short time zero crossing rate is studied. The experiment shows that the new double threshold method can solve the problem that the traditional double threshold method can not detect the speech segment with low energy.In the part of feature extraction, the feature parameters of the combination of Mel cepstrum number and differential Mel frequency cepstrum coefficient are adopted, which better reflect the speaker's personality characteristics.Then, the Gao Si mixed model is studied, and the initialization method of the model parameters is proposed, which combines split method and K-means clustering method, and then two endpoint detection algorithms are proposed by the Gao Si mixed model.The simulation results of feature extraction algorithm and training method show that the system has good recognition effect in pure speech environment.One of the difficulties in the research of voiceprint recognition is the research of recognition system under background noise.Although the performance of the recognition system in pure speech environment is very good, the recognition rate is obviously decreased in the noise environment.In this paper, wavelet transform and wavelet packet transform are used to deal with noise. The denoising effect of wavelet packet is obviously better than that of wavelet transform. Then, an improved threshold denoising method is proposed on the basis of common threshold and compromise threshold of wavelet packet.Through the comparison and simulation of speech signal and the experimental data of the whole system, it is shown that the improved threshold algorithm based on wavelet packet can remove the noise very well, and the recognition system after denoising has achieved a high recognition rate.Finally, the algorithm is applied to complex noise processing, and the algorithm is still effective in removing noise.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【引证文献】