基于嵌入式声纹识别系统的研究与实现

发布时间：2018-03-28 10:45

本文选题：声纹识别　切入点：特征提取　出处：《广东工业大学》2012年硕士论文

【摘要】：近年来,声纹识别技术在逐渐的成熟,声纹识别作为一种生物认证技术,有其独特的优点,如声音是非接触式的,自然的,用户容易接受。因为语言这一媒介的优势,通过语音身份认证技术,声纹识别迅速应用到实际,突出了巨大的市场潜力,声纹识别技术已成为一个新兴的高技术产业。随着计算机硬件和软件技术,半导体技术,电子技术,通信技术和网络技术的发展,以及嵌入式技术的不断发展和更新,其性能和便携性大大提高。实时数据采集,滤波处理,可以在低功耗,体积小的嵌入式设备完成。今天,处理器因为其特殊的结构和高的编译效率使其能够快速的实现声纹识别算法,满足今天的数字信号处理和高实时性的要求。高性能嵌入式声纹识别系统的声纹识别技术,因为方便,经济性,准确性和嵌入式系统的便携性,移动性等优点,被广泛应用于人们的日常生活,拥有广阔的发展前景。本文在分析声纹识别的相关理论与技术的基础上,重点研究了基于Mel倒谱系数(MFCC)的特征参数的提取和DTW算法进行改进,对一些不足之处进行相应的改进。最后,它被应用在基于ARM11与WinCE嵌入式平台下实现的一个小容量的嵌入式声纹识别系统。在前人工作的基础上,本文改进工作主要包括以下三个方面： 1.特征提取方面：对标准的MFCC中存在的不足,提出了相应的改进,加权差分结合MFCC语音特征参数。使用短时帧能量和短时加权过零率替代MFCC中有负识别作用的第1,2阶分量,并根据语音成分的不同贡献率进行加权,然后进行一阶差分,最终会合并成一个新的特征参数。 2.DTW算法方面：使用改进的DTW算法,替代标准的DTW算法,采用整体路径约束,该算法具有很好的鲁棒性,从而提高了算法的效率和代码质量。 3.嵌入式系统实现方面：在基于ok6410的arm11嵌入式系统中的资源相对有限的条件下,进行了一些优化处理。包括操作系统的优化定制和移植,通过跨平台的软件开发,成功在搭建好的嵌入式开发平台上实现了声纹识别系统。并研究分析了改进的DTW算法和传统DTW算法之间的性能差异,对在嵌入式中的运行情况进行了分析。该系统相关的实验,实验结果表明,对同一文本的内容,识别系统的识别率比较高,对文本无关的内容,识别率应该改进；用改进后的算法和特征参数,系统的平均识别率提高4%左右。
[Abstract]:In recent years, voicerecognition technology has gradually matured. As a biometric authentication technology, voicerecognition has its unique advantages, such as sound is contactless, natural and easy to accept by users, because of the advantage of language as a medium. Through the voice identification technology, voicerecognition is applied to practice rapidly, which highlights the huge market potential. Voicerecognition technology has become a new high-tech industry. With the computer hardware and software technology, semiconductor technology, With the development of electronic technology, communication technology and network technology, as well as the continuous development and update of embedded technology, its performance and portability are greatly improved. Today, because of its special structure and high compilation efficiency, the processor can quickly realize the voiceprint recognition algorithm. To meet the requirements of today's digital signal processing and high real-time. High performance embedded voice recognition system voiceprint recognition technology, because of the advantages of convenience, economy, accuracy and embedded system portability, mobility, and other advantages, Widely used in people's daily life, has a broad development prospects. On the basis of analyzing the theory and technology of voiceprint recognition, this paper focuses on the feature parameter extraction based on Mel cepstrum coefficient and the improvement of DTW algorithm. It is applied to a small capacity embedded voiceprint recognition system based on ARM11 and WinCE embedded platform. Based on the previous work, the improvement work in this paper mainly includes the following three aspects:. 1. Feature extraction: for the shortcomings of standard MFCC, a corresponding improvement is put forward. The weighted difference is combined with MFCC speech feature parameters. The second order component with negative recognition in MFCC is replaced by short-time frame energy and short-time weighted zero-crossing rate. The speech components are weighted according to different contribution rates, and then the first order difference is carried out, which will be merged into a new feature parameter. In the aspect of 2.DTW algorithm, the improved DTW algorithm is used instead of the standard DTW algorithm and the global path constraint is adopted. The algorithm has good robustness and improves the efficiency and code quality of the algorithm. 3. The realization of embedded system: under the condition of limited resources in arm11 embedded system based on ok6410, some optimization processes are carried out, including the optimized customization and transplantation of operating system, and the development of cross-platform software. The voiceprint recognition system is successfully implemented on a well built embedded development platform, and the performance difference between the improved DTW algorithm and the traditional DTW algorithm is analyzed, and the running situation in the embedded system is analyzed. The experimental results show that the recognition rate of the system is high for the content of the same text, the recognition rate should be improved for the text-independent content, and the improved algorithm and feature parameters should be used. The average recognition rate of the system is increased by about 4%.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP368.1;TN912.34

【引证文献】