基于盲源分离的车载语音增强算法研究

发布时间：2018-10-20 19:26

【摘要】：语音作为一种方便、快捷、有效的交流方式,在人们的日常生活中扮演着非常重要的角色。随着社会科技的不断进步及其人工智能的迅猛发展,语音信号也逐渐成为人-机交互的一种重要方式,其较传统的人-机交互方式更加的便捷、高效和安全,故被广泛应用于工业控制、医疗辅助、安防保障、智能家居等诸多方面。然而在实际的应用场景中语音信号不可避免的会受到周围环境噪声的干扰,进而影响语音质量,导致其无法完成正常的人-机交互功能。因此语音增强作为一个能够有效抑制噪声分量,提高语音质量的方法,具有重要的研究意义和应用价值。针对车载环境这一特定的应用场景,噪声信号具有低频分布、先验知识不易获得、与语音信号混合情况复杂等特点,造成了许多语音增强算法并不能很好的适用于车载环境。因此本文在分析车载噪声和车载声学场景的基础上,建立噪声信号和语音信号的卷积混合模型,研究盲源分离(Blind Source Separation,BSS)技术在车载环境下进行语音增强的有效性和可行性,以提高车载环境下带噪语音信号的质量和可懂度。本文具体开展了以下的工作:(1)车载声学场景分析建模和噪声估计算法研究。根据车载环境所固有的特点,分析车载噪声的来源及其和驾驶员语音信号在车内的传播路径,建立噪声信号和语音信号在车内的卷积混合模型。由于多数语音增强算法都需要噪声的估计值作为消噪的先验知识,因此噪声估计的准确性将直接影响这些语音增强算法的性能。本文在归纳总结一些常用的语音处理理论基础上,对现有常用的噪声估计算法进行了研究,包括语音端点检测噪声估计算法和最小值控制递归平均噪声估计算法。(2)语音质量评价和语音增强算法研究。文章归纳总结了一些常用的语音信号质量主客观评价标准,并分析了这些评价标准的优缺点。同时针对真实环境下客观评价标准缺少参考源这一问题,本文构建了一个基于隐马尔科夫模型(Hidden Markov Model,HMM)的小词汇量语音识别引擎,并将语音识别率纳入了无参考源语音质量的评价体系中。对于语音增强算法的研究,文章首先实验性分析了谱减法和维纳滤波法这两个经典的语音增强算法,并给出了它们对车载带噪语音信号的消噪结果;其次针对一些传统语音增强算法的不足,本文提出了一个改进的小波阈值函数语音增强算法,该算法可有效抑制宽带噪声和提高语音质量;最后文章阐述了独立分量分析(Independent Component Analysis,ICA)的基本理论框架和实现原理,并重点研究了利用基于负熵的复值ICA在频域盲解卷积,实现语音增强的过程。该ICA语音增强过程不仅可以较好的契合卷积混合模型,而且可以很好的弥补现有语音增强算法在车载环境中应用的不足。(3)基于卷积ICA的车载语音增强算法研究。文章根据语音信号和车载噪声信号的卷积混合特性以及它们在频域的非高斯分布特性,提出利用基于负熵极大的卷积ICA对车载带噪语音信号进行语音增强,并对该增强过程进行针对性的优化。文章在仿真环境,室内环境,真实车载环境三种声学场景下构建了车载带噪语音信号语料库,并采用基于负熵的卷积ICA进行语音消噪。实验结果表明,该卷积ICA消噪后语音信号的识别率较车载带噪语音信号分别最高提高了18.33%,30%,27.5%,展现出该卷积ICA在车载声学场景中应用的有效性和鲁棒性。最后本文针对频域盲解卷积ICA的语音消噪效果受语音信号分帧长度和帧移大小影响的问题进行了实验性研究和阐述。(4)复杂环境下语音增强系统的研究和实现。本文在所研究的噪声估计算法和语音增强算法基础上,选择部分算法结合语音媒体控制逻辑,在Windows平台下利用C++实现了一套复杂环境下的语音增强系统。该系统具有语音波形显示,频谱显示,选择性语音增强、语音播放保存等功能。测试结果表明,该系统不但具有较好的语音增强性能,同时获得了较强的可靠性和鲁棒性。
[Abstract]:As a convenient, fast and effective way of communication, speech plays a very important role in people's daily life. Along with the progress of social science and technology and the rapid development of artificial intelligence, the voice signal gradually becomes an important way of human-machine interaction. It is more convenient, efficient and safe than the traditional man-machine interactive mode, so it is widely used in industrial control and medical assistance. Security and security, smart home and other aspects. However, in the actual application scene, the voice signal is inevitably disturbed by surrounding environmental noise, and then the voice quality is influenced, and the normal person-machine interaction function can not be completed. Therefore, speech enhancement plays an important role in suppressing noise components and improving the quality of speech. Aiming at this particular application scene of vehicle-mounted environment, the noise signal has low frequency distribution, the prior knowledge is not easy to obtain, and the mixing condition of the voice signal is complicated and the like, so that many voice enhancement algorithms do not apply to the vehicle-mounted environment very well. Therefore, based on the analysis of vehicle-mounted noise and vehicle-mounted acoustic scene, this paper establishes the convolution mixture model of noise signal and speech signal, and researches the validity and feasibility of blind source separation (BSS) technology in vehicle-mounted environment. so as to improve the quality and the intelligibility of the noisy speech signal under the vehicle-mounted environment. In this paper, the following work is carried out: (1) on-board acoustic scene analysis modeling and noise estimation algorithm research. According to the inherent characteristics of vehicle-mounted environment, the source of vehicle-mounted noise and the propagation path of the driver's voice signal in the vehicle are analyzed, and the convolution mixture model of the noise signal and the voice signal in the vehicle is established. Since most speech enhancement algorithms require an estimate of noise as a priori knowledge of noise cancellation, the accuracy of the noise estimation will directly affect the performance of these speech enhancement algorithms. On the basis of summarizing some common speech processing theories, this paper studies the existing commonly used noise estimation algorithms, including the speech endpoint detection noise estimation algorithm and the minimum control recursive average noise estimation algorithm. (2) Speech quality evaluation and speech enhancement algorithm research. The main objective evaluation criteria of speech signal quality are summarized in this paper, and the advantages and disadvantages of these evaluation standards are analyzed. At the same time, we construct a small vocabulary speech recognition engine based on Hidden Markov Model (HMM) based on Hidden Markov Model (HMM) and integrate the speech recognition rate into the evaluation system without reference source speech quality. For the research of speech enhancement algorithm, the paper firstly analyzes the two classical speech enhancement algorithms of spectral subtraction and Wiener filtering, and gives their noise elimination results for vehicle-mounted noisy speech signal; secondly, aiming at the deficiency of some traditional speech enhancement algorithms, This paper presents an improved speech enhancement algorithm for small wave threshold functions, which can effectively suppress wideband noise and improve speech quality. Finally, the basic theoretical framework and implementation principle of Independent Component Analysis (ICA) are described. In this paper, we focus on the process of using complex value ICA based on negative entropy in the frequency domain blind deconvolution to realize the speech enhancement. The ICA speech enhancement process not only can better fit the convolution mixing model, but also can make up the deficiency of the existing speech enhancement algorithm in the vehicle-mounted environment. (3) Research on vehicle-mounted speech enhancement algorithm based on convolution ICA. Based on the convolution mixing characteristics of speech signal and vehicle-mounted noise signal and their non-Gaussian distribution in frequency domain, the speech enhancement of vehicle-mounted noisy speech signal using convolution ICA based on negative entropy is proposed, and the enhancement process is optimized. In this paper, an on-board noisy speech signal corpus was constructed under three acoustic scenes of environment, indoor environment and real vehicle environment, and the speech noise was eliminated by convolution ICA based on negative entropy. The experimental results show that the recognition rate of the speech signal after the convolution ICA is improved by 18. 33%, 30% and 27. 5% respectively, which shows the validity and robustness of the convolution ICA in the vehicle-mounted acoustic scene. In the end, the speech noise elimination effect of blind deconvolution ICA in frequency domain is studied and explained by the influence of frame length and frame shift size of speech signal. (4) Research and implementation of speech enhancement system under complex environment. In this paper, based on the studied noise estimation algorithm and speech enhancement algorithm, a part of the algorithm is selected to combine with the speech media control logic, and the speech enhancement system under a complex environment is realized with C ++ under the Windows platform. The system has the functions of voice waveform display, frequency spectrum display, selective speech enhancement, voice play preservation and the like. The test results show that the system not only has better speech enhancement performance, but also has better reliability and robustness.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】