数字助听器声信号处理关键算法研究

发布时间：2018-08-12 15:18

【摘要】：随着老龄化社会的到来,老龄听损患者的数量持续增加。听损对老年人精神上造成的困扰仅次于位居第一的消化系统疾病,长期患有老年性听损会严重影响老年人的身心健康并降低其生活质量。对于轻度、中度甚至重度感音性听损患者来说,配戴助听器是最有效的听力干预和康复手段。近一个世纪以来,助听器技术取得了长足进步,用户对助听器的满意度从1980年的59.6%上升至2008年的74.0%。但是,现有助听器的性能因受到啸叫、信号畸变以及噪声的影响而难以提高。部分患者已经验配了助听器,但是没有佩戴,其中一个重要的原因就是助听器不能提供足够的益处。因此,面向听损患者的数字助听器技术亟需进一步研究。本文从响度补偿、自适应声反馈抑制、非线性频率压缩算法和语音可懂度增强等方面对汉语数字助听器的几种关键技术进行了研究,主要研究内容如下:1、论文提出了一种自适应响度补偿方法,用于缓解传统的响度补偿算法中信号可听和畸变之间的矛盾。首先,利用gammatone滤波器组对信号进行多通道分解来模拟人耳听觉特性;然后,根据患者的听觉动态范围和通道信号的声压级来确定补偿方法。若采用非线性补偿方法,则在补偿过程中使用自适应宽动态范围压缩(Wide Dynamic Range Compression,WDRC)技术,该技术根据患者的听觉范围和输出信号的声压级自适应地调整压缩比,使其尽量接近于1。自适应调整压缩比的目的是在信号可听的前提下,减小信号的畸变。并且,自适应WDRC技术可以提高噪声环境下的输出信噪比(Signalto Noise Ratio,SNR)。实验结果表明,与WDRC补偿方法相比,所提方法在语噪声环境下提高了输出SNR;与线性和WDRC补偿方法相比,提出的方法提高了语音可懂度和语音质量。2、论文提出了一种应用于助听器声反馈抑制的变步长算法,用于缓解自适应滤波算法在快收敛速度和低失调量之间的矛盾。算法根据自适应滤波器系数归一化的短时平均和长时平均之间的距离,将滤波器的更新状态分为收敛状态、过渡状态和稳定状态,并在不同的自适应状态下使用不同的步长。在收敛状态使用大步长保证收敛速度;在过渡状态使用阶梯状步长,以进一步降低失调量;在稳定状态使用小步长,以保证系统收敛至稳态解。该变步长算法分别被应用于归一化的最小均方误差(Normalized Least Mean Error,NLMS)自适应滤波器和归一化的子带自适应滤波器(Normalized Sub-band Adaptive Filter,NSAF)算法,形成了变步长 NLMS(Variable Step Size NLMS,VSS-NLMS)和变步长NSAF(Variable Step Siz eNSAF,VSS-NSAF)算法。实验结果表明,本文提出的算法具有收敛速度快,稳态失调量低,输出语音的语音感知质量(Perceptual Evaluation of Speech Quality,PESQ)评价值高等特点。3、论文提出了一种自适应非线性频率压缩算法,以充分利用听损患者残余的可听频带,提高高频听损患者的语音可懂度。首先,对高于截止频率的信号按照BARK尺度进行临界带分割;其次,根据患者的截止频率和最大可听频率确定全局压缩比;再次,根据全局压缩比和子带信号归一化的平均能量自适应地确定子带压缩比,从而确定子带信号压缩后的频带范围;最后,通过映射的方法将高频信号压缩至低频,使得高频不可听的信号变得可听。语音可懂度实验结果表明,与WDRC算法和非线性频率压缩算法相比,本文提出的算法可显著提高声母以及不同信噪比条件下句子的可懂度。4、论文提出了一种新的语音特征——多分辨率功率归一化倒谱系数(Multi-Resolution Power Normalized Cepstram Coefficients,MRPNCC)特征来获取语音的局部信息和上下文信息。并且,基于MRPNCC特征,提出了一种语音可懂度增强算法,以提高听损患者的语音可懂度。首先,利用MRPNCC特征向量训练支撑矢量机(Support Vector Machine,SVM)模型;其次,利用训练好的模型对语音的时频单元进行分类,移除以噪声为主导的时频单元,保留以语音为主导的时频单元,并使用维纳滤波器对其进行去噪;最后,将滤波后的时频信号重构得到增强的语音。实验结果表明,MRPNCC特征优于传统的谱特征,而且语音可懂度实验结果表明,所提出的语音可懂度增强算法可以显著提高听损患者的语音可懂度。
[Abstract]:With the advent of an aging society, the number of elderly patients with hearing loss continues to increase. The mental distress of the elderly is second only to that of the digestive system. Chronic suffering from hearing loss in the elderly will seriously affect their physical and mental health and reduce their quality of life. For mild, moderate and even severe sensorineural hearing loss patients Hearing aids are the most effective means of hearing intervention and rehabilitation. In the past century, the technology of hearing aids has made great progress. The satisfaction of users with hearing aids has increased from 59.6% in 1980 to 74.0% in 2008. However, the performance of existing hearing aids is difficult to improve due to the effects of howling, signal distortion and noise. Hearing AIDS have been tested but not worn. One of the important reasons is that they do not provide sufficient benefits. Therefore, digital hearing aids for hearing impaired patients need to be further studied. Several key techniques of Chinese digital hearing aids are studied. The main research contents are as follows: 1. An adaptive loudness compensation method is proposed to alleviate the contradiction between the signal audibility and distortion in the traditional loudness compensation algorithm. Firstly, the Gammatone filter banks are used to decompose the signals into multi-channels to simulate human hearing. The compensation method is then determined according to the patient's auditory dynamic range and the sound pressure level of the channel signal. If the nonlinear compensation method is used, the adaptive wide dynamic range compression (WDRC) technique is used in the compensation process, which is based on the patient's auditory range and the sound pressure level of the output signal. Adaptive compression ratio can be adjusted as close as possible to 1. The purpose of adaptive compression ratio is to reduce signal distortion while the signal is audible. Moreover, adaptive WDRC technology can improve the output signal to Noise Ratio (SNR) in noisy environments. Experimental results show that the proposed method is more effective than WDRC compensation method in terms of noise. Compared with linear and WDRC compensation methods, the proposed method improves speech intelligibility and speech quality. 2. A variable step size algorithm for hearing aids acoustic feedback suppression is proposed to alleviate the contradiction between fast convergence speed and low offset. The distance between the normalized short-term average and the long-term average of the filter coefficients divides the update state of the filter into convergence state, transition state and stable state, and uses different step sizes in different adaptive states. The variable step size algorithm is applied to the normalized Least Mean Error (NLMS) adaptive filter and the normalized subband adaptive filter (NSAF) algorithm respectively to form the variable step size algorithm. Variable Step Size NLMS (VSS-NLMS) and Variable Step Size NSAF (VSS-NSAF) algorithms are proposed. Experimental results show that the proposed algorithm has the advantages of fast convergence, low steady-state offset, and high Perceptual Evaluation of Speech Quality (PESQ) of output speech. An adaptive nonlinear frequency compression algorithm is proposed to make full use of the residual audible frequency band of the patients with hearing loss and improve the speech intelligibility of the patients with high frequency hearing loss. According to the global compression ratio and the normalized average energy of the sub-band signal, the sub-band compression ratio is adaptively determined to determine the frequency range of the compressed sub-band signal. Finally, the high-frequency signal is compressed to low-frequency by mapping method, which makes the high-frequency non-audible signal audible. Compared with the linear frequency compression algorithm, the proposed algorithm can significantly improve the intelligibility of consonants and sentences under different SNR conditions. 4. A new speech feature, multi-resolution power Normalized Cepstram Coefficients (MRPNCC), is proposed to obtain speech localization. Furthermore, a speech intelligibility enhancement algorithm based on MRPNCC features is proposed to improve the speech intelligibility of hearing impaired patients. Firstly, the support vector machine (SVM) model is trained by using MRPNCC feature vectors; secondly, the time-frequency units of speech are classified and removed by using the trained model. The noise-dominated time-frequency unit retains the speech-dominated time-frequency unit and uses Wiener filter to denoise it. Finally, the filtered time-frequency signal is reconstructed to enhance the speech. The experimental results show that the MRPNCC features are superior to the traditional spectral features, and the speech intelligibility experimental results show that the proposed speech is intelligible. The degree enhancement algorithm can significantly improve the speech intelligibility of hearing impaired patients.
【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R764.5;TN912.3

【相似文献】