基于语音信号处理的呼吸认证方法研究

发布时间：2018-10-09 08:20

【摘要】：说话人识别是一类基础问题,该问题可以细分为两类:说话人辨别问题和说话人确认问题。根据对语音文本的依赖程度不同,说话人识别又分为文本相关型、文本无关型、文本提示型。识别说话人包括训练/登记和测试/验证两个阶段,训练阶段建立合法说话人(即目标说话人)特征模板,测试阶段则计算待识别说话人的测试语音数据与合法说话人特征模板的相似度,并得出判断结果。在最近20多年的研究中,说话人识别普遍关注语音信号的音源,发音系统,韵律特征。而语音中的呼吸信号在研究中并没有受到关注,甚至将其作为噪音项进行剔除。通过本文的观察和研究发现,呼吸信号具有其独特性,是呼吸系统意义上的生理指纹。它是人体自然而然产生的,持续时间短,发生频率低且处理简单,呼吸特征是人体与生俱来的,具有唯一,稳定的特点。基于呼吸信号的这些特点,本文提出了一种基于呼吸特征的说话人识别方案—“BreathID”,它具有文识别准确率高、文本无关、轻量级且可实现对用户透明的优点。“Breath ID”分为3个部分:呼吸边界分离,特征提取和选择,特征匹配。本文通过CDF统计分析和实证研究,选择梅尔频率倒谱系数(Mel Frequency Cepstral Coefficients,简称MFCC)作为说话人的呼吸特征。文中提出了基于一组简单向量操作的轻量级的分类算法,最后通过该分类算法进行相似度匹配,做出说话人决策。“Breath ID”的实验分为两个部分,第一部分是对50个用户收集数据进行实验,评估“BreathID”的总体性能。实验结果如下:说话人辨别实验中,方案的错误辨别率(False Identification Rate:简称FIR)为0.04%;说话人确认实验中,方案的错误接受率(False Acceptance Rate:简称FAR)为0.12%,错误拒绝率(False Rejection Rate:简称FAR)为0.15%。第二部分通过另外20个用户的数据在不同的实用场景中对“BreathID”的一致性进行评估,包括文本无关性,用户状态(静坐和走路),录音设备(3个智能手机和2个商业麦克风),录音周期(8个月),语言(中文和英文),攻击(重放攻击)。
[Abstract]:Speaker recognition is a basic problem, which can be subdivided into two categories: speaker identification problem and speaker confirmation problem. According to the different degree of dependence on phonetic text, speaker recognition can be divided into text correlation type, text independent type and text prompt type. The identification of the speaker consists of two stages: training / registration and testing / verification. In the training stage, the legal speaker (i.e. the target speaker) feature template is established. In the testing stage, the similarity between the test speech data and the legal speaker feature template is calculated, and the judgment results are obtained. In the past 20 years, speaker recognition has focused on phonetic sources, pronunciation systems and prosodic features of speech signals. However, the respiratory signals in speech have not been paid attention to, and even eliminated as noise terms. It is found that respiratory signals are unique and physiological fingerprints in the sense of respiratory system. It is naturally produced by the human body, with short duration, low frequency of occurrence and simple handling. The breathing characteristics are inherent to the human body and have unique and stable characteristics. Based on these characteristics of respiratory signals, this paper proposes a speaker recognition scheme based on respiratory characteristics, called "BreathID", which has the advantages of high accuracy of text recognition and text independence. "Breath ID" is divided into three parts: respiratory boundary separation, feature extraction and selection, and feature matching. Based on CDF statistical analysis and empirical study, Mel frequency cepstrum coefficient (Mel Frequency Cepstral Coefficients, (MFCC) is chosen as the breathing feature of the speaker. In this paper, a lightweight classification algorithm based on a set of simple vector operations is proposed. Finally, the classification algorithm is used to match the similarity and make the speaker decision. The experiment of "Breath ID" is divided into two parts. The first part is an experiment of 50 users collecting data to evaluate the overall performance of BreathID. The experimental results are as follows: in the speaker discrimination experiment, the error discrimination rate (False Identification Rate:) of the scheme is 0.04, the error acceptance rate (FAR) of the scheme is 0.12 and the error rejection rate (FAR) is 0.15. The second part evaluates the consistency of "BreathID" in different practical scenarios using data from another 20 users, including text independence. User status (sit-in and walk), recording device (3 smartphones and 2 commercial microphones), recording cycle (8 months), language (Chinese and English), attack (replay attack).
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】