时频字典学习的单通道语音增强算法

针对以往语音增强算法在非平稳噪声环境下性能急剧下降的问题,基于时频字典学习方法提出了一种新的单通道语音增强算法。首先,提出采用时频字典学习方法对噪声的频谱结构的先验信息进行建模,并将其融入到卷积非负矩阵分解的框架下;然后,在固定噪声时频字典情况下,推导了时变增益和语音时频字典的乘性迭代求解公式;最后,利用该迭代公式更新语音和噪声的时变增益系数以及语音的时频字典,通过语音时频字典和时变增益的卷积运算重构出语音的幅度谱并用二值时频掩蔽方法消除噪声干扰。实验结果表明,在多项语音质量评价指标上,本文算法都取得了更好的结果。在非平稳噪声和低信噪比环境下,相比于多带谱减法和非负稀疏编码去噪算法,本文算法更有效地消除了噪声,增强后的语音具有更好的质量。     

一种增强带噪语音可懂度的新算法

一般的语音增强算法在强噪声环境中只能提高信噪比,不能提高可懂度。本文提出用可调节白噪声代替信号中非语音部分的语音可懂度增强处理新算法。实验证明此方法能明显改善强噪声时的语音可懂度,能对低至-10dB的带噪语音信号进行有效的可懂度增强。     

提高耳语音可懂度的非对称压缩语音增强方法

提出两种基于非对称代价函数的耳语音增强算法,将语音增强过程中的放大失真和压缩失真区分对待。Modified ItakuraSaito (MIS)算法对放大失真给予更多的惩罚,而Kullback-Leibler (KL)算法则对压缩失真给予更多的惩罚。实验结果表明,在低于—6 dB的低信噪比情况中,经MIS算法增强后的耳语音的可懂度相比传统算法有显著提高;而KL算法则获得了同最小均方误差语音增强算法近似的可懂度提高效果,证实了耳语音中的放大失真和压缩失真对于耳语音可懂度的影响并不相同,低信噪比时较大的压缩失真有助于提高耳语音可懂度,而高信噪比时的压缩失真对耳语音可懂度影响较小。     

联合深度神经网络和凸优化的单通道语音增强算法

噪声估计的准确性直接影响语音增强算法的好坏,为提升当前语音增强算法的噪声抑制效果,有效求解无约束优化问题,提出一种联合深度神经网络(DNN)和凸优化的时频掩蔽优化算法进行单通道语音增强。首先,提取带噪语音的能量谱作为DNN的输入特征;接着,将噪声与带噪语音的频带内互相关系数(ICC Factor)作为DNN的训练目标;然后,利用DNN模型得到的互相关系数构造凸优化的目标函数;最后,联合DNN和凸优化,利用新混合共轭梯度法迭代处理初始掩蔽,通过新的掩蔽合成增强语音。仿真实验表明,在不同背景噪声的低信噪比下,相比改进前,新的掩蔽使增强语音获得了更好的对数谱距离(LSD)、主观语音质量(PESQ)、短时客观可懂度(STOI)和分段信噪比(segSNR)指标,提升了语音的整体质量并且可以有效抑制噪声。     

面向自定义语音唤醒的关键词相关的单通道语音增强

提出一种面向自定义语音唤醒的单通道语音增强方法。该方法预先将关键词音素信息存入文本编码矩阵,并在常规语音增强模型基础上添加一个基于注意力机制的音素偏置模块。该模块利用语音增强模型中间特征从文本编码矩阵中获取当前帧的音素信息,并将其融入语音增强模型的后续计算中,从而提升语音增强模型对关键词相关音素的增强效果。在不同噪声环境下的实验结果表明,该方法可以更有效地抑制关键词部分噪声。同时所提出方法对比常规语音增强方法与其他文本相关语音增强方法,在自定义语音唤醒性能上可以分别获得14.3%和7.6%的相对提升。     

基于字典学习和稀疏表示的单通道语音增强算法综述*

如何从带噪语音信号中恢复出干净的语音信号一直都是信号处理领域的热点问题。近年来研究者相继提出了一些基于字典学习和稀疏表示的单通道语音增强算法,这些算法利用语音信号在时频域上的稀疏特性,通过学习训练数据样本的结构特征和规律来构造相应的字典,再对带噪语音信号进行投影以估计出干净语音信号。针对训练样本与测试数据不匹配的情况,有监督类的非负矩阵分解方法与基于统计模型的传统语音增强方法相结合,在增强阶段对语音字典和噪声字典进行更新,从而估计出干净语音信号。本文首先介绍了单通道情况下语音增强的信号模型,然后对4种典型的增强方法进行了阐述,最后对未来可能的研究热点进行了展望。     

氦语音增强算法的研究

当潜水员深海作业时，由于生理方面的原因，需要以氦氧混合气体为呼吸介质替代空气，但这时会出现所谓的“氦语音”现象，语音产生很大畸变，清晰度极低。本文给出了氦语音的基本变化规律，介绍了两种分别基于短时傅立叶变换和线性预测模型的增强算法。我们的实验结果表明，两种算法均能显著校正氦语音，提高清晰度。     

稀疏低秩噪声模型下无监督实时单通道语音增强算法

针对现有基于字典学习的增强算法需要先验信息、不易实时处理的问题,提出一种便于实时处理的无监督的单通道语音增强算法。首先,该算法将无监督条件下背景噪声的建模问题转化为带噪语音幅度谱的稀疏低秩噪声分解;然后,采用增量非负子空间方法对背景噪声进行在线字典学习,获得能够体现背景噪声时变特性的自适应噪声字典;最后,利用所得的噪声字典,采用易于实时处理的逐帧迭代方式,对带噪语音进行处理。实验结果表明:相较于多带谱减法和基于低秩稀疏矩阵分解的增强算法,所提算法在噪声抑制方面的性能尤为显著,在多项性能评价指标上,均表现出更好的结果。     

稀疏低秩噪声模型下无监督实时单通道语音增强算法

针对现有基于字典学习的增强算法需要先验信息、不易实时处理的问题,提出一种便于实时处理的无监督的单通道语音增强算法。首先,该算法将无监督条件下背景噪声的建模问题转化为带噪语音幅度谱的稀疏低秩噪声分解;然后,采用增量非负子空间方法对背景噪声进行在线字典学习,获得能够体现背景噪声时变特性的自适应噪声字典;最后,利用所得的噪声字典,采用易于实时处理的逐帧迭代方式,对带噪语音进行处理。实验结果表明:相较于多带谱减法和基于低秩稀疏矩阵分解的增强算法,所提算法在噪声抑制方面的性能尤为显著,在多项性能评价指标上,均表现出更好的结果。     

汉语音段反转言语的可懂度研究

实验研究了帧长对汉语音段反转言语可懂度的影响。实验结果表明,帧长在64 ms以下,汉语音段反转言语具有较高的可懂度;帧长在64~203 ms之间,可懂度随帧长的增加逐渐降低;帧长在203 ms以上,可懂度为0。在帧长8 ms时,汉语的声调失真导致可懂度下降。原始语音信号和音段反转言语的调制谱的分析表明,调制谱失真大小和可懂度密切相关。因此,用原始语音信号和音段反转言语的窄带包络间的归一化相关值可以衡量调制谱失真大小,基于语音的语言传输指数法计算的客观值和实验结果显著相关(r=0.876,p<0.01)。研究表明,语言可懂度与窄带包络有关,音段反转言语的可懂度和保留原始语音信号的窄带包络密切相关。     

Comparative intelligibility investigation of single-channel noise-reduction algorithms for Chinese, Japanese, and English

A large number of single-channel noise-reduction algorithms have been proposed based largely on mathematical principles. Most of these algorithms, however, have been evaluated with English speech. Given the different perceptual cues used by native listeners of different languages including tonal languages, it is of interest to examine whether there are any language effects when the same noise-reduction algorithm is used to process noisy speech in different languages. A comparative evaluation and investigation is taken in this study of various single-channel noise-reduction algorithms applied to noisy speech taken from three languages: Chinese, Japanese, and English. Clean speech signals (Chinese words and Japanese words) were first corrupted by three types of noise at two signal-to-noise ratios and then processed by five single-channel noise-reduction algorithms. The processed signals were finally presented to normal-hearing listeners for recognition. Intelligibility evaluation showed that the majority of noise-reduction algorithms did not improve speech intelligibility. Consistent with a previous study with the English language, the Wiener filtering algorithm produced small, but statistically significant, improvements in intelligibility for car and white noise conditions. Significant differences between the performances of noise-reduction algorithms across the three languages were observed.     

Effects of hearing protector devices on speech intelligibility

The purpose of this study was to determine the influence of hearing protection devices (HPDs) on the understanding of speech in young adults with normal hearing, both in a silent situation and in the presence of ambient noise. The experimental research was carried out with the following variables: five different conditions of HPD use (without protectors, with two earplugs and with two earmuffs); a type of noise (pink noise); 4 test levels (60, 70, 80 and 90 dB[A]); 6 signal/noise ratios (without noise, +5, +10, zero, −5 and −10 dB); 5 repetitions for each case, totalling 600 tests with 10 monosyllables in each one. The variable measure was the percentage of correctly heard words (monosyllabic) in the test. The results revealed that, at the lowest levels (60 and 70 dB), the protectors reduced the intelligibility of speech (compared to the tests without protectors) while, in the presence of ambient noise levels of 80 and 90 dB and unfavourable signal/noise ratios (0, −5 and −10 dB), the HPDs improved the intelligibility. A comparison of the effectiveness of earplugs versus earmuffs showed that the former offer greater efficiency in respect to the recognition of speech, providing a 30% improvement over situations in which no protection is used. As might be expected, this study confirmed that the protectors' influence on speech intelligibility is related directly to the spectral curve of the protector's attenuation.     

人耳听觉相关代价函数深度学习单通道语声增强算法

均方误差函数是深度学习单通道语声增强算法最常用的一种代价函数。然而,均方误差值的大小与语声质量好坏并非完全相关。为了提高算法性能,该文在深度神经网络训练中引入了两类与人耳听觉相关的代价函数。第一类是加权欧氏距离代价函数,考虑了人耳听觉掩蔽效应;第二类是Itakura-Satio代价函数、COSH代价函数和加权似然比代价函数,强调语声谱峰的重要性,侧重于恢复干净语声谱峰信息。基于长短期记忆网络结构分析比较了两类代价函数在深度学习单通道语声增强算法中的性能,并与均方误差代价函数进行对比。实验结果表明,基于加权欧式距离代价函数的深度神经网络单通道语声增强算法能够获得更好的语声质量和更低的噪声残留。     

Effect of the speed of a single-channel dynamic range compressor on intelligibility in a competing speech task

Using a "noise-vocoder" cochlear implant simulator [Shannon et al., Science 270, 303-304 (1995)], the effect of the speed of dynamic range compression on speech intelligibility was assessed, using normal-hearing subjects. The target speech had a level 5 dB above that of the competing speech. Initially, baseline performance was measured with no compression active, using between 4 and 16 processing channels. Then, performance was measured using a fast-acting compressor and a slow-acting compressor, each operating prior to the vocoder simulation. The fast system produced significant gain variation over syllabic timescales. The slow system produced significant gain variation only over the timescale of sentences. With no compression active, about six channels were necessary to achieve 50% correct identification of words in sentences. Sixteen channels produced near-maximum performance. Slow-acting compression produced no significant degradation relative to the baseline. However, fast-acting compression consistently reduced performance relative to that for the baseline, over a wide range of performance levels. It is suggested that fast-acting compression degrades performance for two reasons: (1) because it introduces correlated fluctuations in amplitude in different frequency bands, which tends to produce perceptual fusion of the target and background sounds and (2) because it reduces amplitude modulation depth and intensity contrasts.     

Effects of low-pass filtering on intelligibility of periodically interrupted speech

The combined effect of low-pass filtering (cut-off frequencies between 500 and 3000 Hz) and periodic interruptions (1.5 and 10 Hz) on speech intelligibility was investigated. When combined, intelligibility was lower than each manipulation alone, even in some conditions where there was no effect from a single manipulation (such as the fast interruption rate of 10 Hz). By using young normal-hearing listeners, potential suprathreshold deficits and aging effects that may occur due to hearing impairment were eliminated. Thus, the results imply that reduced audibility of high-frequency speech components may partially explain the reduced intelligibility of interrupted speech in hearing impaired persons.     

Effects of temporal and spectral factors of maskers on speech intelligibility

This study demonstrates a new possibility of estimating intelligibility of speech in informational maskers. The temporal and spectral properties of sound maskers are investigated to achieve acoustic privacy in public spaces. Speech intelligibility (SI) tests were conducted using Japanese sentences in daily use for energy (white noise) or informational (reversed speech) maskers. We found that the masking effects including informational masking on SI might not be estimated by analyzing the narrow-band temporal envelopes, which is a common way of predicting SI under noisy conditions. The masking effects might instead be visualized by spectral auto-correlation analysis on a frame-by-frame basis, for the series of dominant-spectral peaks of the masked target in the frequency domain. Consequently, we found that dissimilarity in frame-based spectral-auto-correlation sequences between the original and masked targets was the key to evaluating maskers including informational masking effects on SI.     

Gain-induced speech distortions and the absence of intelligibility benefit with existing noise-reduction algorithms

Most noise-reduction algorithms used in hearing aids apply a gain to the noisy envelopes to reduce noise interference. The present study assesses the impact of two types of speech distortion introduced by noise-suppressive gain functions: amplification distortion occurring when the amplitude of the target signal is over-estimated, and attenuation distortion occurring when the target amplitude is under-estimated. Sentences corrupted by steady noise and competing talker were processed through a noise-reduction algorithm and synthesized to contain either amplification distortion, attenuation distortion or both. The attenuation distortion was found to have a minimal effect on speech intelligibility. In fact, substantial improvements (>80 percentage points) in intelligibility, relative to noise-corrupted speech, were obtained when the processed sentences contained only attenuation distortion. When the amplification distortion was limited to be smaller than 6 dB, performance was nearly unaffected in the steady-noise conditions, but was severely degraded in the competing-talker conditions. Overall, the present data suggest that one reason that existing algorithms do not improve speech intelligibility is because they allow amplification distortions in excess of 6 dB. These distortions are shown in this study to be always associated with masker-dominated envelopes and should thus be eliminated.