基于多基音跟踪的单声道混合语音分离*

针对许多计算听觉场景分析系统无法很好地解决多说话人混合语音信号分离的问题,提出了一种基于多基音跟踪的单声道混合语音分离系统。该系统充分利用了多基音跟踪研究的最新成果,通过将多基音跟踪得到的目标语音和干扰语音的基音轨迹信息结合到分离系统中,有效地改善了分离系统在包括多说话人混合在内的多种干扰情况下的分离效果,为多说话人语音分离问题的解决提供了新的思路。     

基于聚类的双说话人混合语音分离

针对许多基于训练模型的计算机听觉场景分析系统,在解决双说话人混合语音信号分离时需要依赖样本训练的有效性以及说话人的先验知识,提出一种基于聚类的单声道混合语音分离系统。系统先利用多基音跟踪算法对语音信号进行分析并产生同时流,然后通过最大化类内散布矩阵与类间散布矩阵的迹,搜索同时流的最佳分类,最终完成对双说话人的语音分离。该系统不需要训练语音模型,并且有效地改善了在双说话人混合语音信号的分离效果,为双说话人的语音分离提供了新的思路。     

基于计算听觉场景分析的混合语音信号分离算法研究

人耳听觉系统能够在强噪声的环境下区分出自己感兴趣的语音,基于计算听觉场景分析(CASA)的基本原理,其重点和难点是找到合适的声音分离线索,完成目标语音信号和噪声信号的分离.针对单通道浊语音分离的问题,提出了一种以基音为线索的浊语音信号分离算法.在白噪声、鸡尾酒会噪声等六种噪声干扰条件下,通过仿真实验结果表明,相比于传统的谱减法,语音分离算法的输出信噪比平均提高了7.47 dB,并有效抑制了干扰噪声,改善了分离效果.     

多个声源下基于人耳听觉特性的语音分离

受声学研究启发,结合人脑人耳听觉特性对语音的处理方式,建立了一个完整的模拟听觉中枢系统的语音分离模型.首先利用外周听觉模型对语音信号进行多频谱分析,然后建立重合神经元模型提取语音信号的特征,最后在脑下丘的神经细胞模型中完成对语音的分离.基于现有的语音识别方法,该模型能够很好地解决绝大多数的语音识别方法都只能在单声源和低噪声的环境下使用的问题.实验结果表明,该模型能够实现多声源环境下语音的分离并且具有较高的鲁棒性.随着研究的深入,基于人耳听觉特性的语音分离模型将有很广泛的应用前景.     

基于计算听觉场景分析的单通道语音分离方法

人耳听觉系统能够从嘈杂的环境中筛选出自己感兴趣的语音,基于计算听觉场景分析的方法,论文采用倒谱法提取语音基音周期轨迹,以连续的基音周期轨迹为线索,按基音频率的整数倍提取各次谐波的频谱,再通过傅里叶逆变换重构分离后的语音.实验表明,在几种典型噪音环境下,该方法能有效将目标语音从背景噪声中分离,信噪比(SNR)和评价意见分...     

3.2 kbps MMBE声码器的研究

提出了一种改进的多带激励（ｍｏｄｉｆｉｅｄ ｍｕｌｔｉ－ｂａｎｄ ｅｘｃｉｔａｔｉｏｎ,ＭＭＢＥ）语音压缩编码算法以适用于低码率通信系统。作为一种多带激励（Ｍｕｌｔｉ－ｂａｄｎ ｅｘｃｉｔａｔｉｏｎ,ＭＢＥ）编码器,ＭＭＢＥ根据重构信号谱与原始信号谱之间的相似程度来进行基音估计和浊清音判决,而其中基音参数的估计准确性将直接影响编码器的性能。文中提出的ＭＭＢＥ算法采用了一种改进的基音估计算法以及基间     

基于计算听觉场景分析的单声道浊音分离

针对单声道语音分离中浊音分离的问题,提出了一种准确估计基音周期的方法。首先,以语音的短时平稳性和基音周期的连续性等为线索,利用语音信号的倒谱峰值构成基音周期谱图,并自动提取基音周期轨迹。然后,利用谐波频率为基音频率整数倍的性质来拾取各次谐波的频谱。最后,通过傅里叶逆变换对浊音进行重构。实验结果表明,该方法能准确提取基音周期轨迹,有效分离浊音信号。     

同态解卷处理在基音检测中的应用

语音处理中基音检测是极为重要的环节之一,然而浊音中的基音往往会受到声道特性和噪声的影响而导致检测结果的误差。利用同态解卷处理,将浊音中的激励信号和声道特性进行分离,然后再在激励信号中利用自相关检测基音,可以减小声道特性和噪声对基音检测的影响,从而提高基音检测的精度。通过理论模型验证了该方法的可行性,而且实际语音信号处理结果表明,该方法在基音检测时可以基本不受声道特性和噪声的影响。     

基于双麦克风的室内语音分离与声源定位系统

为了探究利用两个麦克风进行多声源分离和二维平面定位的可能性,提出了一种基于双麦克风的室内语音分离与声源定位系统。该系统根据麦克风采集的信号,建立了双麦克风时延-衰减模型,然后利用DUET算法估计了模型的时延-衰减参数,并绘制了参数直方图。在语音分离阶段,建立了二进制时频掩膜（BTFM）,根据参数直方图,结合二值掩蔽的方法对混合语音进行了分离;在声源定位阶段,通过推导模型衰减参数与信号能量比之间的关系,得到了确定声源位置的数学方程组。利用Roomsimove工具箱模拟室内声学环境,通过Matlab仿真和几何坐标计算,在对多个声源目标分离的同时完成了二维平面中的定位。实验结果表明,该系统对多个声源信号的定位误差均在2%以下,有助于小型系统的研究和开发。     

基于AMDF的藏语语音基音周期检测

基音周期是语音信号最重要的参数之一,它描述了语音激励源的一个重要特征。被广泛应用到语音合成、语音识别等领域。本文介绍了一种基于AMDF的语音基音周期检测方法,较好的提取了语音的基音周期。     

背景噪声下的语音信号分离

独立分量分析法在分离含有背景噪声的混合语音时效果不理想。为此,将独立分量分析算法与卡尔曼滤波相结合,对语音进行降噪处理,采用FastICA算法对含噪语音进行分离,分离速率高于Informax算法,能够获得较清晰的语音文件。通过仿真验证了该方法的可行性和有效性。     

Unsupervised Single-Channel Music Source Separation by Average Harmonic Structure Modeling

Source separation of musical signals is an appealing but difficult problem, especially in the single-channel case. In this paper, an unsupervised single-channel music source separation algorithm based on average harmonic structure modeling is proposed. Under the assumption of playing in narrow pitch ranges, different harmonic instrumental sources in a piece of music often have different but stable harmonic structures; thus, sources can be characterized uniquely by harmonic structure models. Given the number of instrumental sources, the proposed algorithm learns these models directly from the mixed signal by clustering the harmonic structures extracted from different frames. The corresponding sources are then extracted from the mixed signal using the models. Experiments on several mixed signals, including synthesized instrumental sources, real instrumental sources, and singing voices, show that this algorithm outperforms the general nonnegative matrix factorization (NMF)-based source separation algorithm, and yields good subjective listening quality. As a side effect, this algorithm estimates the pitches of the harmonic instrumental sources. The number of concurrent sounds in each frame is also computed, which is a difficult task for general multipitch estimation (MPE) algorithms.     

Formula: a programming language for expressive computer music

An overview is given of Formula (an abbreviation for Forth Music Language), a language for controlling synthesizers that can model the expressiveness of a human performance. Formula supports algorithmic composition, interactive performance, and programmed interpretation of traditional scores. It uses concurrent processes that share a single address space and are scheduled by the runtime system. Note-playing processes compute sequences of pitches and play these pitches as notes or chords. Auxiliary processes are attached to note-playing processes or groups to supply note parameters such as volume, duration, and articulation. Input-handling processes execute when input arrives from a particular device. Two representative Formula programs are described     

基于计算听觉场景分析的语音混叠信号分离

基于计算听觉场景原理，提出了一种混叠语音信号分离算法模型，对两个说话者的混叠声音进行分离。该模型对低频区和高频区的分离分别采用了不同方法，避免了因采用同样方法处理低频高频区而导致对高频段语音不能很好分离的结果。实验结果表明，该模型具有很好的应用效果。     

基于卷积非负矩阵部分联合分解的强噪声单声道语音分离

非负矩阵部分联合分解(Nonnegative matrix partial co-factorization, NMPCF)将指定源频谱作为边信息参与混合信号频谱的联合分解, 以帮助确定指定源的基向量进而提高信号分离性能.卷积非负矩阵分解(Convolutive nonnegative matrix factorization, CNMF)采用卷积基分解的方法进行矩阵分解, 在单声道语音分离方面取得较好的效果.为了实现强噪声条件下的语音分离, 本文结合以上两种算法的优势, 提出一种基于卷积非负矩阵部分联合分解(Convolutive nonnegative partial matrix co-factorization, CNMPCF)的单声道语音分离算法.本算法首先通过基音检测算法得到混合信号的语音起始点, 再据此确定混合信号中的纯噪声段, 最后将混合信号频谱和噪声频谱进行卷积非负矩阵部分联合分解, 得到语音基矩阵, 进而得到分离的语音频谱和时域信号.实验中, 混合语音信噪比(Signal noise ratio, SNR)选择以-3 dB为间隔从0 dB至-12 dB共5种SNR.实验结果表明, 在不同噪声类型和噪声强度条件下, 本文提出的CNMPCF方法相比于以上两种方法均有不同程度的提高.     

Maximum likelihood‐based gradient estimation for multivariable nonlinear systems using the multiinnovation identification theory

This article considers the identification problems of multivariable input nonlinear systems with unmeasured disturbances. For the identification difficulty caused by the crossproducts between the parameters of the linear block and the nonlinear block, the key term separation technique is adopted to separate the parameters of the nonlinear block from the parameters of the linear block. By combining the model decomposition technique and the hierarchical identification principle, a key term separation‐based maximum likelihood recursive extended stochastic gradient algorithm with reduced computational complexity is presented to estimate all the parameters directly. By introducing the multiinnovation identification theory, a key term separation‐based maximum likelihood multiinnovation extended stochastic gradient algorithm is proposed to improve the parameter estimation accuracy. The simulation results illustrate the effectiveness of the proposed methods.     

Creating High Confidence in a Separation Kernel

Separation of processes is the foundation for security and safety properties of systems. This paper reports on a collaborative effort of Government, Industry and Academia to achieve high confidence in the separation of processes. To this end, this paper will discuss (1) what a separation kernel is, (2) why the separation of processes is fundamental to security systems, (3) how high confidence in the separation property of the kernel was obtained, and (4) some of the ways government, industry, and academia cooperated to achieve high confidence in a separation kernel. What is separation? Strict separation is the inability of one process to interfere with another. In a separation kernel, the word separation is interpreted very strictly. Any means for one process to disturb another, be it by communication primitives, by sharing of data, or by subtle uses of kernel primitives not intended for communication, is ruled out when twoprocesses are separated. Why is separation fundamental? Strict separation between processes enables the evaluation of a system to check that the system meets its security policy. For example, if a red process is strictly separated from a black process, then it can be concluded that there is no flow of information from red to black. How was high confidence achieved? We have collaborated and shared our expertise in the use of SPECWARE. SPECWARE is a correct by construction method, in which high level specifications are built up from modules using specification combinators. Refinements of the specifications are made until an implementation is achieved. These refinements are also subject to combinators. The high confidence in the separation property of the kernel stems from the use of formal methods in the development of the kernel. How did we collaborate? Staff from the Kestrel Institute (developers of SPECWARE), the Department of Defense (DoD), and Motorola (developers of the kernel) cooperated in the creation of the Mathematically Analyzed Separation Kernel (MASK). DoD provided the separation kernel concept, and expertise in computer security and high confidence development. Kestrel provided expertise in SPECWARE. Motorola combined its own the expertise with that of DoD and Kestrel in creating MASK.     

Estimation of wiener nonlinear systems with measurement noises utilizing correlation analysis and Kalman filter

This paper is concerned with parameter estimation of Wiener systems with measurement noises employing correlation analysis method and adaptive Kalman filter. The presented Wiener system consists of two series blocks, that is, a dynamic block represented by auto-regressive moving average (ARMA) model, and static nonlinear block established by neural fuzzy model. Aim at estimating separately the two blocks, the separable signals are introduced. First, applying the separable signals to decouple the identification of linear dynamic block from that of static nonlinear block, then ARMA model parameters are estimated employing correlation function-based least squares principle. Moreover, aiming at handle with error caused by colored measurement noise, adaptive Kalman filter technique and cluster method are introduced to estimate parameter of the nonlinear block and noises model, enhancing parameter estimation precision. The accuracy and applicability of estimated scheme presented are verified through numerical simulation and nonlinear process, the results demonstrate that it is feasible for estimating the Wiener systems in the presence of colored measurement noises.     

Hammerstein system identification through best linear approximation inversion and regularisation

Hammerstein systems are composed by the cascading of a static nonlinearity and a linear system. In this paper, a methodology for identifying such systems using a combination of least squares support vector machines (LS-SVM) and best linear approximation (BLA) techniques is proposed. To do this, a novel method for estimating the intermediate variable is presented allowing a clear separation of the identification steps. First, an approximation to the linear block is obtained through the BLA of the system. Then, an approximation to the intermediate variable is obtained using the inversion of the estimated linear block and the known output. Afterwards, a nonlinear model is calculated through LS-SVM using the estimated intermediate variable and the known input. To do this, the regularisation capabilities of LS-SVM play a crucial role. Finally, a parametric re-estimation of the linear block is made. The method was tested in three examples, two of them with hard nonlinearities, and was compared with four other methods showing very good performance in all cases. The obtained results demonstrate that also in the presence of noise, the method can effectively identify Hammerstein systems. The relevance of these findings lies in the fact that it is shown how the regularisation allows to bypass the usual problems associated with the noise backpropagation when the inversion of the estimated linear block is used to compute the intermediate variable.     

Correction of pathological speeches and assistance to learners with vocal disabilities

This work describes a new methodology for correcting voice defects contained in the Arabic speeches and assisting learners of Arabic vocabulary. For this purpose, we follow four stages. The first step consists in localizing the vocal disabilities which degrade an Arabic voice signal, so we focus on comparing between a referenced probabilistic-phonetic model and a speaker model. Second, we differentiate two cases: Degraded speeches can be generated from pathological problems, or it can be produced by non arabophone learners. Hence, we compare between forced alignment scores. Third, we develop a new algorithm to correct pathological pronunciations. The last task is the conception of an application assisting learners of Arabic vocabulary in improving their pronunciation. The achieved results are encouraging. Moreover, learners of Arabic vocabulary have presented a good amelioration using the developed application. A lot of applications that design systems of voice signal processing can use our proposition.