文本无关的多说话人确认研究

发布时间：2018-08-27 09:52

【摘要】：近年来,在生物特征识别领域,说话人识别以其独特的安全性、经济性和准确性等优势受到越来越多的关注,并逐渐成为人们生活和工作中重要的身份验证方式,具有广阔的市场前景。说话人识别的一个重要研究分支是说话人确认,本文着重对说话人确认展开研究。本文从说话人确认的系统框架入手,对系统的各部分予以详细的介绍。随后针对复杂条件下的说话人确认问题,重点研究了特征提取、说话人分割、模型建立等技术。本文的主要研究工作及创新点如下:1.构建基于GMM-UBM的说话人确认系统并将其作为本文的基线系统,研究分析了影响系统性能的相关因素,包括高斯混合度、训练语音长度、得分规整技术,并通过实验进行验证。2.在特征提取方面,为了提升噪声环境下说话人确认系统的性能,本文提出了一种具有较强噪声鲁棒性的多窗谱减MFCC特征。多窗谱减MFCC是在已有多窗谱MFCC(Multitaper MFCC)基础上的改进,主要是将多窗谱估计技术与谱减法进行了结合。仿真结果表明,当测试语音中含有加性噪声时,与多窗谱MFCC提取算法相比,采用多窗谱减MFCC的说话人确认系统性能在等错误率EER和最小检测代价函数值minDCF两项评测指标上都取得了较好的结果。3.在说话人分割方面,针对传统基于BIC的说话人分割算法累积计算量大、冗余分割点过多,导致分割速度慢、分割准确度降低的问题,相关文献采用了分治算法对其进行改进,虽然改进法能够大幅提高分割速度,但准确度却有所降低。为了达到分割速度与分割准确度同时提高的目的,本文首先在具体实现BIC说话人分割算法时提出了三步分割的策略,在此基础上引入分治算法思想对其进行改进。实验结果表明,改进后的分割算法在分割速度上有较大提高,准确度上也有一定提升。4.在模型建立方面,探索研究了i-vector说话人建模技术,重点研究了i-vector的提取过程,构建基于i-vector的说话人确认系统,并将其与基于GMM-UBM的说话人确认系统进行了对比分析。
[Abstract]:In recent years, in the field of biometrics, speaker recognition has attracted more and more attention because of its unique advantages of security, economy and accuracy, and has gradually become an important way of identity verification in people's lives and work. It has broad market prospects. This paper begins with the system framework of speaker verification, and then introduces each part of the system in detail. Then, aiming at the speaker verification under complex conditions, it focuses on feature extraction, speaker segmentation, model building and other technologies. The main research work and innovation of this paper are as follows: 1. Based on the GMM-UBM speaker verification system as the baseline system of this paper, the related factors affecting the performance of the system are studied and analyzed, including Gaussian mixture, training speech length, scoring regularization technology, and verified by experiments. 2. In feature extraction, in order to improve the performance of speaker verification system in noisy environment, this paper proposes a method to improve the performance of the system. A multi-window spectral subtraction MFCC feature with strong noise robustness is proposed. The multi-window spectral subtraction MFCC is an improvement on the existing multi-window spectral MFCC (Multitaper MFCC), which combines the multi-window spectral estimation technique with the spectral subtraction method. The simulation results show that when the test speech contains additive noise, it is better than the multi-window spectral MFCC extraction algorithm. The speaker verification system using multi-window spectral subtraction MFCC achieves good results in EER with equal error rate and minDCF with minimum detection cost function. In order to improve the segmentation speed and accuracy at the same time, this paper first proposes a three-step segmentation strategy to implement the BIC speaker segmentation algorithm. The experimental results show that the improved segmentation algorithm has a great improvement in segmentation speed and accuracy. 4. In the aspect of model building, I-vector speaker modeling technology is explored and studied, especially the extraction process of I-vector and the construction of I-vector based speaker. The speaker recognition system is analyzed and compared with the speaker verification system based on GMM-UBM.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【共引文献】