Informed acoustic source separation and localization

In this paper, localization and separation of acoustic sources are examined. Depending on the number of sources in relation to the array channels, the problem is investigated in terms of underdetermined and overdetermined configurations. In the underdetermined configuration, virtual monopole sources are assumed in uniformly spaced angles. The problem is then formulated into compressive sampling (CS) problem which can be solved by using the linearly constrained -norm convex (CVX) optimization. The solution yields the directions of real sources and the source signal spectrum, which enables localization and reconstruction of sources at one shot. In the underdetermined configuration, source localization and signal separation is carried out in two steps. First, the directions of arrival (DOA) are estimated with Minimum Variance Distortionless Response (MVDR) or Multiple Signal Classification (MUSIC). Next, Tikhonov regularization (TIKR) is utilized to recover the source spectrum. In the localization problem for both configurations, Neyman-Pearson detector is employed to determine thresholds for source detection. Numerical and experimental results show that the proposed methods produce improved speech quality in terms of mean opinion score (MOS) in perceptual evaluation of speech quality (PESQ) test.

Blind single-channel source separation is a long-standing machine learning and signal processing problem. Traditional blind source separation (BSS) algorithms were proposed to solve this task utilizing multiple signal constraints. Generative adversarial network (GAN) are free from statistical constraints and samples, but the role of adversarial training in the task of BSS has not been fully demonstrated. Therefore, a new separation network model that enables to learn the known separated signal distribution from the stepwise fine estimation of the unknown mixture distribution was presented in this paper, and a self-attention mechanism was introduced to solve the problem of the blurring details of the generated image by the generator which preserves image details in the process of image separation. Compared with the existing single-channel blind source separation algorithm based on generative adversarial network-neural egg separation (NES), the detailed information of this new separation algorithm is more prominent, and the source signal in the mixed image has been separated more effectively, and has better separation performance than the classic blind source separation algorithms.

To separate and localize the P300 sources within the brain, a robust constrained blind source separation (CBSS) algorithm has been proposed. The algorithm is an extension of the Infomax algorithm, based on minimization of mutual information, for which a reference P300 signal is used as a constraint. The reference signal forces the unmixing matrix to separate the sources of both auditory and visual P300 resulting from the corresponding stimulations. The constrained problem is then converted to an unconstrained problem by means of a set of nonlinear penalty functions. The P300 sources are then localized based on electrode source inner products.

To separate and localize the P300 sources within the brain a robust constrained blind source separation (CBSS) algorithm has been proposed. The algorithm is an extension of the Infomax algorithm, based on minimization of mutual information, for which a reference P300 signal is used as a constraint. The reference signal forces the unmixing matrix to separate the sources of both auditory and visual P300 resulted from the corresponding stimulations. The constrained problem is then converted to an unconstrained problem by means of a set of nonlinear penalty functions. The P300 sources arc then localized based on electrode source correlations.

In the recent past, blind source separation (BSS) algorithms using multivariate statistical data analysis technique have been successfully used for source identification and separation in the field of biomedical and statistical signal processing. Recently numbers of different BSS techniques have been developed. With BSS methods being the feasible method for source separation and decomposition of biosignals, it is important to compare the different techniques and determine the most suitable method for the applications. This paper presents the performance of five BSS algorithms (SOBI, TDSEP, FastICA, JADE and Infomax) for decomposition of sEMG to identify subtle finger movements. It is observed that BSS algorithms based on second-order statistics (SOBI and TDSEP) gives better performance compared to algorithms based on higher-order statistics (FastICA, JADE and infomax).

Reverberant speech separation with probabilistic time–frequency masking for B-format recordings

We propose a dynamic latent variable model for learning latent bases from time varying, non-negative data. We take a probabilistic approach to modeling the temporal dependence in data by introducing a dynamic Dirichlet prior—a Dirichlet distribution with dynamic parameters. This new distribution allows us to assure non-negativity and avoid intractability when sequential updates are performed otherwise encountered in using Dirichlet prior. We refer to the proposed model as the Dirichlet latent variable model DLVM. We develop an expectation maximization algorithm for the proposed model, and also derive a maximum a posteriori estimate of the parameters. Furthermore, we connect the proposed DLVM to two popular latent basis learning methods—probabilistic latent component analysis PLCA and non-negative matrix factorization NMF. We show that 1 PLCA is a special case of our DLVM, and 2 DLVM can be interpreted as a dynamic version of NMF. The usefulness of DLVM is demonstrated for three audio processing applications—speaker source separation, denoising, and bandwidth expansion. To this end, a new algorithm for source separation is also proposed. Through extensive experiments on benchmark databases, we show that the proposed model outperforms several relevant existing methods in all three applications.

For linear mixing model (LMM) of hyperspectral unmixing in hyperspectral images processing problem, the endmember fractional abundances satisfy the sum-to-one constraint, which makes the well-known independent component analysis (ICA) based blind source separation (BSS) algorithms not well suited to blind hyperspectral unmixing (bHU). In this paper, an efficient non-parametric bHU algorithm consulting dependent component analysis (DCA) is presented. Based on the cumulative density function (CDF) and order statistics instead of traditional probability distribution function (PDF), the novel objective function is derived by maximising the non-parametric non-Gaussianity between the estimated endmember abundance of the endmember signatures and their corresponding original abundances. With the stochastic gradient rule of constrained optimisation method, an efficient dependent sources separation algorithm for bHU is obtained to fulfil the endmember signatures extraction and abundances estimation tasks. Simulations based on the synthetic data are performed to evaluate the validity of the proposed non-parametric non-Gaussianity HU (non-pNG-bHU) algorithm.

A new blind image source separation algorithm based on feedback sparse component analysis

A new brain source localization technique using electroencephalograms (EEGs) is investigated in this paper. The information which describes the location of certain known sources is used as the constraint within the proposed blind source separation (BSS) algorithm and leads to a solution to the ill-posed inverse problem of source localization. Non-homogeneity of the head tissues, on the other hand, is exploited by introducing a realistic model of the mixing system. This model is used to better identify the location of the unknown sources within the brain from projection of the separated independent components on to the scalp. A separate procedure is employed to highlight the rhythmic EEG sources such as Alpha rhythm as the known sources. The performance of the scheme is shown on real EEG measurements and compared with that of "conventional dipole fitting algorithm".

Source separation (SS) and feature extraction (FE) are tools employed in digital signal processing. The former permits to estimate the values of some sources that have been mixed, and the latter extracts features from a set of measurements. SS and FE are widely applied on biomedical signals such as electrocardiogram (ECG), electroencephalogram, arterial blood pressure, etc., because these signals are collected in noisy environments. For instance, ECG recordings show the electrical activity generated by the whole heart at once. Yet, there are cardiac pathologies or arrhythmias related to only the atrial or ventricular chambers and thus, tools capable to separate them become fundamental for the diagnosis, prognosis and prediction of life-threatening events. Therefore, the quality of treatments depends on the reliability of the features extracted from the signals and then, the reduction of possible interferences becomes very relevant in this context. The study and the development of new SS technique play an important role in those application in which the components of a measurement cannot be splitted using classical temporal or frequency analysis. In addition, non-blind SS aims to employ further information and to develop mathematical and statistical model to make the estimates more reliable. Features extraction is fundamental for the classification task. In biomedical signals, features are used to characterize the status of the subjects in either healthy or pathological condition. For example, features to predict the risk of developing cardiac arrhythmias are continuously encouraged by regulatory agencies as the US Food and Drug Administration. The study of reliability and feasibility of features requires an extensive use of tests. These tests are necessary to evaluate some properties, e.g., the capability of the feature to be resilient to noise, variability of the estimate, classification power, etc. The aim of this thesis is to study, develop, validate and test new SS techniques and features applicable to different kind of signals. The new algorithms and features are extensively studied to characterize their properties from a methodological point of view. In addition, simulated and real data are considered as a test bench. Cardiac signals will be the specific field of application. First, a new algorithm for non-blind SS will be presented and discussed. In particular, this new methodology is an extension of a well-known algorithm, i.e., template, matching and subtraction (TMS), normally used to estimate transient sources, i.e., sources that are located only somewhere in the signals, in stationary conditions. TMS estimates the values of the source by averaging a set of measurements in which it is known to be constant over time. However, there are situations in which this assumption does not hold and the results obtained are not “good” estimates. In order to track changes over

Most source separation algorithms are implemented as spectrogram decomposition. In contrast, time-domain source separation is less investigated since there is a lack of an efficient signal representation that facilitates decomposing oscillatory components of a signal directly in the time domain. In this letter, we utilize the Ramanujan subspace and the nested periodic subspace to address this issue, by constructing a parametric dictionary that emphasizes period information with less redundancy. Methods including iterative subspace projection and convolutional sparse coding can decompose a mixture into signals with distinct oscillation periods according to the dictionary. Experiments on score-informed source separation show that the proposed method is competitive to the state-of-the-art, frequency-domain approaches when the provided pitch information and the signal parameters are the same.

The task of manipulating the level and/or effects of individual instruments to recompose a mixture of recordings, or remixing, is common across a variety of applications such as music production, audio-visual post-production, podcasts, and more. This process, however, traditionally requires access to individual source recordings, restricting the creative process. To work around this, source separation algorithms can separate a mixture into its respective components. Then, a user can adjust their levels and mix them back together. This two-step approach, however, still suffers from audible artifacts and motivates further work. In this work, we re-purpose Conv-TasNet, a well-known source separation model, into two neural remixing architectures that learn to remix directly rather than just to separate sources. We use an explicit loss term that directly measures remix quality and jointly optimize it with a separation loss. We evaluate our methods using the Slakh and MUSDB18 datasets and report remixing performance as well as the impact on source separation as a byproduct. Our results suggest that learning-to-remix significantly outperforms a strong separation baseline and is particularly useful for small volume changes.

Advances in machine learning (ML) and the internet-of-things (IoT) have resulted in a renewed interest in analog matrix-vector multiplication (MvM) accelerators [1–3]. Classification based tasks have exploited low-to-medium resolution multiplication and accuracy boosting algorithms in order to compensate for the reduced resolution. Complementing classification, tasks like source separation and localization have diverse applications ranging from signal conditioning in communication [4] and ultrasound to electroencephalography (EEG) [5] source localization and spike sorting, and greatly benefit from similar algorithms. However, due to their lower resolution and limited channel count previously developed systems cannot be directly applied to this task. High-resolution analog multiplication introduces challenges that have limited prior work to less than 6-bit multiplication in the analog domain. Alternative approaches utilizing very high oversampling result in very inefficient solutions. High precision in matrix-multiplication can mitigate the effects of ill-conditioned (almost singular) matrices, as with beamforming separation of near-collinear sources, and other tasks incurring principal component analysis [3] or independent component analysis (ICA) [5]. As seen in Fig. 21.7.1, a large signal dynamic range at the input can results in an untenable dynamic range specification on the downstream data-converters leading to greater than 10× increase in power [4]. Thus, we present a multichannel multiple-input multiple-out (MIMO) mixed-signal linear transform system, with analog signal path and digital coefficient control, composed of an array of 14-bit Nested Thermometer Multiplying DACs (NTMDACs) implementing analog multiplication, and variable gain amplifier (VGA) implementing accumulation. We demonstrate state-of-the art performance on two tasks, spectrally oblivious interference suppression in communication signals and EEG signal separation.

A frequently encountered problem in signal processing is harmonic retrieval in additive colored Gaussian or non-Gaussian noise, especially when the frequencies of the harmonic signals are very close in space. The purpose of this paper is to develop an efficient Blind Source Separation (BSS) algorithm from linear mixtures of source signals, which enables to separate harmonic source signals using only one observed channel signal even if the frequencies of the harmonic signals are closely spaced. First, we establish the BSS based harmonic retrieval model in additive noise by using the only one observed channel, and analyze the fundamental principle by utilizing BSS method to retrieve harmonics. Then, we propose a BSS-based approach to the harmonic retrieval by resorting the concept of W-disjoint orthogonality in the over-complete BSS situation, and as a result, we get the separation algorithm using only one channel mixed signals. Simulation results show that the proposed separation algorithm-BSS-HR is able to separate the harmonic source signals.