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。      

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。     

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。      

一种抑制方向性噪声的双耳语音增强算法

为了给双耳听力设备佩戴者带来更好的语音可懂度,提出了一种利用双耳时间差与声级差的近场语音增强算法,该方法首先利用这两种差异来估计语音的功率谱和语音的相干函数,然后计算干扰噪声在左右耳间的头相关传输函数的比值,最后构造两个维纳滤波器。客观评价的参数显示该算法去噪效果优于对比算法而目标语音的时间差误差和声级差误差低于对比算法。主观的言语接受阈测试表明该方法能有效提高语音可懂度。结果表明,该算法在能够有效去除干扰噪声的同时,保留了目标语音的空间信息。      

改进的正交匹配追踪的语音增强算法*

为了提高传统正交匹配追踪（Orthogonal Matching Pursuit,OMP ）算法的语音增强性能和运算速度,本研究基于稀疏编码理论,提出了一种改进的OMP算法的语音增强算法。其一,将K-奇异值分解（K-singular value decomposition,K-SVD）算法与OMP算法相结合,通过设置能量阈值的方法,提高OMP算法的语音增强性能;其二,通过改进传统OMP算法中信号稀疏逼近的计算方法,提高算法的运算速度。改进的OMP算法的语音增强算法与传统K-SVD语音增强算法相比,采用PESQ评价增强语音的质量,NCM评价语音的可懂度。在NCM的值基本保持不变的情况下,PESQ的值平均提高约12.47%,取得了更好的增强效果。取得了更好的增强效果。改进的OMP算法的运算速度与传统OMP算法相比提高近一倍。     

基于听觉掩蔽效应和Bark子波变换的语音增强

提出了一种适用于低信噪比下的提高语音的听觉效果的语音增强方法。该方法在谱减法的基础上有两个特点:首先减参数是根据人耳听觉掩蔽效应提出的且是自适应的;其次采用了与人耳听觉系统特性更为适应的Bark子波变换方法对增强前后的语音进行分析。对该算法进行了客观和主观测试,结果表明:与谱减法相比对低信噪比的语音信号,(1)能更好地抑制残留噪声和背景噪声,(2)增强后的语音具有更好的清晰度和可懂度。     

U-net网络中融合多头注意力机制的单通道语音增强EI北大核心CSCD

在低信噪比和突发背景噪声条件下,已有的深度学习网络模型在单通道语音增强方面效果并不理想,而人类可以利用语音的长时相关性对不同的语音信号形成综合感知。因此刻画语音的长时依赖关系有助于改进低信噪比和突发背景噪声下的增强性能。受该特性的启发,提出一种融合多头注意力机制和U-net深度网络的增强模型TU-net,实现基于时域的端到端单通道语音增强。TU-net网络模型采用U-net网络的编解码层对带噪语音信号进行多尺度特征融合,并利用多头注意力机制实现双路径Transformer,用于计算语音掩模,更好地建模长时相关性。该模型在时域、时频域和感知域计算损失函数,并通过加权组合损失函数指导训练。仿真实验结果表明,TU-net在低信噪比和突发背景噪声条件下增强语音信号的语音质量感知评估(PESQ)、短时客观可懂度(STOI)和信噪比增益等多个评价指标都优于同类的单通道增强网络模型,且保持相对较少的网络模型参数。     

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。      

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。      

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。     

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.     

Quantifying the intelligibility of speech in noise for non-native listeners

When listening to languages learned at a later age, speech intelligibility is generally lower than when listening to one's native language. The main purpose of this study is to quantify speech intelligibility in noise for specific populations of non-native listeners, only broadly addressing the underlying perceptual and linguistic processing. An easy method is sought to extend these quantitative findings to other listener populations. Dutch subjects listening to Germans and English speech, ranging from reasonable to excellent proficiency in these languages, were found to require a 1-7 dB better speech-to-noise ratio to obtain 50% sentence intelligibility than native listeners. Also, the psychometric function for sentence recognition in noise was found to be shallower for non-native than for native listeners (worst-case slope around the 50% point of 7.5%/dB, compared to 12.6%/dB for native listeners). Differences between native and non-native speech intelligibility are largely predicted by linguistic entropy estimates as derived from a letter guessing task. Less effective use of context effects (especially semantic redundancy) explains the reduced speech intelligibility for non-native listeners. While measuring speech intelligibility for many different populations of listeners (languages, linguistic experience) may be prohibitively time consuming, obtaining predictions of non-native intelligibility from linguistic entropy may help to extend the results of this study to other listener populations.     

The interlanguage speech intelligibility benefit

This study investigated how native language background influences the intelligibility of speech by non-native talkers for non-native listeners from either the same or a different native language background as the talker. Native talkers of Chinese (n = 2), Korean (n = 2), and English (n = 1) were recorded reading simple English sentences. Native listeners of English (n = 21), Chinese (n = 21), Korean (n = 10), and a mixed group from various native language backgrounds (n = 12) then performed a sentence recognition task with the recordings from the five talkers. Results showed that for native English listeners, the native English talker was most intelligible. However, for non-native listeners, speech from a relatively high proficiency non-native talker from the same native language background was as intelligible as speech from a native talker, giving rise to the "matched interlanguage speech intelligibility benefit." Furthermore, this interlanguage intelligibility benefit extended to the situation where the non-native talker and listeners came from different language backgrounds, giving rise to the "mismatched interlanguage speech intelligibility benefit." These findings shed light on the nature of the talker-listener interaction during speech communication.     

The clear speech effect for non-native listeners

Previous work has established that naturally produced clear speech is more intelligible than conversational speech for adult hearing-impaired listeners and normal-hearing listeners under degraded listening conditions. The major goal of the present study was to investigate the extent to which naturally produced clear speech is an effective intelligibility enhancement strategy for non-native listeners. Thirty-two non-native and 32 native listeners were presented with naturally produced English sentences. Factors that varied were speaking style (conversational versus clear), signal-to-noise ratio (-4 versus -8 dB) and talker (one male versus one female). Results showed that while native listeners derived a substantial benefit from naturally produced clear speech (an improvement of about 16 rau units on a keyword-correct count), non-native listeners exhibited only a small clear speech effect (an improvement of only 5 rau units). This relatively small clear speech effect for non-native listeners is interpreted as a consequence of the fact that clear speech is essentially native-listener oriented, and therefore is only beneficial to listeners with extensive experience with the sound structure of the target language.     

A comparative intelligibility study of single-microphone noise reduction algorithms

The evaluation of intelligibility of noise reduction algorithms is reported. IEEE sentences and consonants were corrupted by four types of noise including babble, car, street and train at two signal-to-noise ratio levels (0 and 5 dB), and then processed by eight speech enhancement methods encompassing four classes of algorithms: spectral subtractive, sub-space, statistical model based and Wiener-type algorithms. The enhanced speech was presented to normal-hearing listeners for identification. With the exception of a single noise condition, no algorithm produced significant improvements in speech intelligibility. Information transmission analysis of the consonant confusion matrices indicated that no algorithm improved significantly the place feature score, significantly, which is critically important for speech recognition. The algorithms which were found in previous studies to perform the best in terms of overall quality, were not the same algorithms that performed the best in terms of speech intelligibility. The subspace algorithm, for instance, was previously found to perform the worst in terms of overall quality, but performed well in the present study in terms of preserving speech intelligibility. Overall, the analysis of consonant confusion matrices suggests that in order for noise reduction algorithms to improve speech intelligibility, they need to improve the place and manner feature scores.     

A simulation study of harmonics regeneration in noise reduction for electric and acoustic stimulation

Recent research results show that combined electric and acoustic stimulation (EAS) significantly improves speech recognition in noise, and it is generally established that access to the improved F0 representation of target speech, along with the glimpse cues, provide the EAS benefits. Under noisy listening conditions, noise signals degrade these important cues by introducing undesired temporal-frequency components and corrupting harmonics structure. In this study, the potential of combining noise reduction and harmonics regeneration techniques was investigated to further improve speech intelligibility in noise by providing improved beneficial cues for EAS. Three hypotheses were tested: (1) noise reduction methods can improve speech intelligibility in noise for EAS; (2) harmonics regeneration after noise reduction can further improve speech intelligibility in noise for EAS; and (3) harmonics sideband constraints in frequency domain (or equivalently, amplitude modulation in temporal domain), even deterministic ones, can provide additional benefits. Test results demonstrate that combining noise reduction and harmonics regeneration can significantly improve speech recognition in noise for EAS, and it is also beneficial to preserve the harmonics sidebands under adverse listening conditions. This finding warrants further work into the development of algorithms that regenerate harmonics and the related sidebands for EAS processing under noisy conditions.     

Speech intelligibility and localization in a multi-source environment.

Natural environments typically contain sound sources other than the source of interest that may interfere with the ability of listeners to extract information about the primary source. Studies of speech intelligibility and localization by normal-hearing listeners in the presence of competing speech are reported on in this work. One, two or three competing sentences [IEEE Trans. Audio Electroacoust. 17(3), 225-246 (1969)] were presented from various locations in the horizontal plane in several spatial configurations relative to a target sentence. Target and competing sentences were spoken by the same male talker and at the same level. All experiments were conducted both in an actual sound field and in a virtual sound field. In the virtual sound field, both binaural and monaural conditions were tested. In the speech intelligibility experiment, there were significant improvements in performance when the target and competing sentences were spatially separated. Performance was similar in the actual sound-field and virtual sound-field binaural listening conditions for speech intelligibility. Although most of these improvements are evident monaurally when using the better ear, binaural listening was necessary for large improvements in some situations. In the localization experiment, target source identification was measured in a seven-alternative absolute identification paradigm with the same competing sentence configurations as for the speech study. Performance in the localization experiment was significantly better in the actual sound-field than in the virtual sound-field binaural listening conditions. Under binaural conditions, localization performance was very good, even in the presence of three competing sentences. Under monaural conditions, performance was much worse. For the localization experiment, there was no significant effect of the number or configuration of the competing sentences tested. For these experiments, the performance in the speech intelligibility experiment was not limited by localization ability.     

The acoustic and perceptual effects of two noise-suppression algorithms

Internal noise generated by hearing-aid circuits can be audible and objectionable to aid users, and may lead to the rejection of hearing aids. Two expansion algorithms were developed to suppress internal noise below a threshold level. The multiple-channel algorithm's expansion thresholds followed the 55-dB SPL long-term average speech spectrum, while the single-channel algorithm suppressed sounds below 45 dBA. With the recommended settings in static conditions, the single-channel algorithm provided lower noise levels, which were perceived as quieter by most normal-hearing participants. However, in dynamic conditions "pumping" noises were more noticeable with the single-channel algorithm. For impaired-hearing listeners fitted with the ADRO amplification strategy, both algorithms maintained speech understanding for words in sentences presented at 55 dB SPL in quiet (99.3% correct). Mean sentence reception thresholds in quiet were 39.4, 40.7, and 41.8 dB SPL without noise suppression, and with the single- and multiple-channel algorithms, respectively. The increase in the sentence reception threshold was statistically significant for the multiple-channel algorithm, but not the single-channel algorithm. Thus, both algorithms suppressed noise without affecting the intelligibility of speech presented at 55 dB SPL, with the single-channel algorithm providing marginally greater noise suppression in static conditions, and the multiple-channel algorithm avoiding pumping noises.     

Second-language experience and speech-in-noise recognition: effects of talker-listener accent similarity

Previous work has shown that the intelligibility of speech in noise is degraded if the speaker and listener differ in accent, in particular when there is a disparity between native (L1) and nonnative (L2) accents. This study investigated how this talker-listener interaction is modulated by L2 experience and accent similarity. L1 Southern British English, L1 French listeners with varying L2 English experience, and French-English bilinguals were tested on the recognition of English sentences mixed in speech-shaped noise that was spoken with a range of accents (French, Korean, Northern Irish, and Southern British English). The results demonstrated clear interactions of accent and experience, with the least experienced French speakers being most accurate with French-accented English, but more experienced listeners being most accurate with L1 Southern British English accents. An acoustic similarity metric was applied to the speech productions of the talkers and the listeners, and significant correlations were obtained between accent similarity and sentence intelligibility for pairs of individuals. Overall, the results suggest that L2 experience affects talker-listener accent interactions, altering both the intelligibility of different accents and the selectivity of accent processing.