Speech Enhancement using K-Sparse Autoencoder Techniques

Effective representation methods and proper signal priors are crucial in most signal processing applications. In this thesis we focus on different structured models and we design appropriate schemes that allow the discovery of low dimensional latent structures that characterise and identify the signals. Motivated by the highly non-linear structure of most datasets, we firstly investigate the geometry of manifolds. Manifolds are low dimensional, non-linear structures that are naturally employed to describe sets of strongly related signals such as the images of a 3-D object captured from different viewpoints. However, the use of manifolds in applications is not straightforward due to their usually non-analytic and non-linear form. We propose here a way to `disassemble' a manifold into simpler components by approximating it with affine subspaces. Our objective is to discover a set of low dimensional affine subspaces that can represent manifold data accurately while preserving the manifold's structure. To this end, we employ a greedy technique that iteratively merges manifold samples into groups based on the difference of local tangents. We use our algorithm to approximate synthetic and real manifolds and to demonstrate that it is competitive to state-of-the-art techniques. Then, we consider structured sparse representations of signals and we propose a new sparsity model, where signals are essentially composed of a small number of structured {\it molecules }. We define the molecules to be linear combinations of a small number of atoms in a redundant dictionary. Our multi-level model takes into account the energy distribution of the significant signal components in addition to their support. It permits to define typical visual patterns and recognise them in prototypical or deformed form. We define a new structural difference measure between molecules and their deformed versions, which is based on their sparse codes and we create an algorithm for decomposing signals into molecules that can account for different deviations in the internal molecule structure. Our experiments verify the benefits of the new image model in various restoration tasks and they confirm that the development of proper models that extend the mere notion of sparsity can be very useful for various inverse problems in imaging. Finally, we investigate the problem of learning molecule representations directly in the sparse code domain. We constrain sparse codes to be linear combinations of a few, possibly deformed, molecules and we design an algorithm that can learn the structure from the codes without transforming them back into the signal domain. To this end, we take advantage of our structural difference which is based on the sparse codes and we devise a scheme for representing the codes with molecules and learn the molecules at the same time. To illustrate the effectiveness of our proposed algorithm we apply it to various synthetic and real datasets and we compare the results with traditional sparse coding and dictionary learning techniques. From the experiments, we verify the superior performance of our scheme in interpreting and recognising correctly the underlying structure. In short, in this thesis we are interested in low-dimensional, structured models. Among the various choices, we focus on manifolds and sparse representations and we propose schemes that enhance their structural properties and highlight their effectiveness in signal representations.

Nonnegative matrix factorization (NMF) is an effective speech separation approach of extracting discriminative components of different speaker. However, traditional NMF focuses only on the additive combination of the components and ignores the dependencies of speeches. Convolutive NMF (CNMF) captures the dependencies of speeches by overlapping components and achieves better separation performance. NMF and CNMF learn dictionaries for speakers in the absence of mixture, and thus they are unable to get enough information to learn dictionaries accurately when testing speeches are available. To handle this problem, transductive NMF (TNMF) is proposed which simultaneously utilizes speech of each speaker and mixture to learn more meaningful features of speakers, and significantly boost speech separation. CNMF addresses the dependencies of speech signals while it ignores the positive effect of mixtures in learning dictionaries. TNMF emphasizes the transductive learning of dictionaries while it fails to consider dependencies of speeches. This paper proposes transductive convolutive NMF (TCNMF) to overcome the deficiencies of both CNMF and TNMF. Experimental results show that our method makes significant improvement compared to aforementioned NMF-based methods.

Multivariate prediction of human behavior from resting state data is gaining increasing popularity in the neuroimaging community, with far-reaching translational implications in neurology and psychiatry. However, the high dimensionality of neuroimaging data increases the risk of overfitting, calling for the use of dimensionality reduction methods to build robust predictive models. In this work, we assess the ability of four well-known dimensionality reduction techniques to extract relevant features from resting state functional connectivity matrices of stroke patients, which are then used to build a predictive model of the associated deficits based on cross-validated regularized regression. In particular, we investigated the prediction ability over different neuropsychological scores referring to language, verbal memory, and spatial memory domains. Principal Component Analysis (PCA) and Independent Component Analysis (ICA) were the two best methods at extracting representative features, followed by Dictionary Learning (DL) and Non-Negative Matrix Factorization (NNMF). Consistent with these results, features extracted by PCA and ICA were found to be the best predictors of the neuropsychological scores across all the considered cognitive domains. For each feature extraction method, we also examined the impact of the regularization method, model complexity (in terms of number of features that entered in the model) and quality of the maps that display predictive edges in the resting state networks. We conclude that PCA-based models, especially when combined with L1 (LASSO) regularization, provide optimal balance between prediction accuracy, model complexity, and interpretability.

Sparse coding or independent component analysis (ICA) which is a holistic representation, was successfully applied to elucidate early auditory processing and to the task of sound classification. In contrast, parts-based representation is an alternative way of understanding object recognition in brain. In this paper we employ the non-negative matrix factorization (NMF) [D.D. Lee et al., 1999] which learns parts-based representation in the task of sound classification. Methods of feature extraction from spectro-temporal sounds using the NMF in the absence or presence of noise are explained. Experimental results show that NMF-based features improve the performance of sound classification over ICA-based features.

By finding broader temporal and spatial patterns of brain activity, dictionary learning and sparse coding techniques circumvent the constraints of typical voxel-level studies of task-based functional magnetic resonance imaging (fMRI) data. Prior applications of these algorithms to task-related fMRI data, on the other hand, have not been tuned to simultaneously uncover temporal patterns of activity that change in tandem with changes in task circumstances and spatial patterns that exploit existing neuroscience knowledge. This work, presents a sparse dictionary learning method that decomposes fMRI data into temporal patterns of signals that loosely differ between task conditions and sparse spatial patterns that are at least partially similar to known functional network hubs, based on prior knowledge of the temporal pattern of task conditions and the locations of brain regions hypothesized to be involved in the task. To over come the a major drawback of the prior work in terms being constrained to provide an equal number of spatial and temporal patterns, even when the data was more accurately represented by a larger number spatial than temporal patterns (or vice versa), this work also propose a constrained decoupled dictionary learning method that uses prior knowledge related to task paradigms and the spatial locations of task-related brain regions to estimate decoupled spatial and temporal patterns that represent the fMRI data. The utility of sparse dictionary learning methods applied to task fMRI data is increased by simultaneously exploiting the task's known temporal structure and biasing solutions towards putative network hubs. The method identifies a high percentage of the spatial and temporal patterns programmed into simulated fMRI data. The temporal and spatial patterns are identified using an effective on-line optimization methodology. The technique detects geographical and temporal patterns in synthetic task fMRI data. When applied to two real-world fMRI data sets, the method: 1. automatically identifies temporal and spatial patterns that were known by neuroscientists to be task-related a priori but were not provided as inputs to the method; 2. provides measurements that differ significantly between differing tasks. When applied to real fMRI data from 100 participants two different studies, the suggested method also identifies geographic regions known a priori to be engaged by the Attention Network Task (ANT), AX-Continuous Performance Task (AX-CPT) and Stroop Task more thoroughly than competing methods. Using task information to partially constrain dictionary learning allowed this method to identify task related networks in real and synthetic fMRI data.

In recent years, sparse coding via dictionary learning has been widely used in many applications for exploiting sparsity patterns of data. For classification, useful sparsity patterns should have discrimination, which cannot be well achieved by standard sparse coding techniques. In this paper, we investigate structured sparse coding for obtaining discriminative class-specific group sparsity patterns in the context of classification. A structured dictionary learning approach for sparse coding is proposed by considering the \(\ell _{2,0}\) norm on each class of data. An efficient numerical algorithm with global convergence is developed for solving the related challenging \(\ell _{2,0}\) minimization problem. The learned dictionary is decomposed into class-specific dictionaries for the classification that is done according to the minimum reconstruction error among all the classes. For evaluation, the proposed method was applied to classifying both the synthetic data and real-world data. The experiments show the competitive performance of the proposed method in comparison with several existing discriminative sparse coding methods.

We propose a framework for automated detection of dry age-related macular degeneration (AMD) and diabetic macular edema (DME) from retina optical coherence tomography (OCT) images, based on sparse coding and dictionary learning. The study aims to improve the classification performance of state-of-the-art methods. First, our method presents a general approach to automatically align and crop retina regions; then it obtains global representations of images by using sparse coding and a spatial pyramid; finally, a multiclass linear support vector machine classifier is employed for classification. We apply two datasets for validating our algorithm: Duke spectral domain OCT (SD-OCT) dataset, consisting of volumetric scans acquired from 45 subjects—15 normal subjects, 15 AMD patients, and 15 DME patients; and clinical SD-OCT dataset, consisting of 678 OCT retina scans acquired from clinics in Beijing—168, 297, and 213 OCT images for AMD, DME, and normal retinas, respectively. For the former dataset, our classifier correctly identifies 100%, 100%, and 93.33% of the volumes with DME, AMD, and normal subjects, respectively, and thus performs much better than the conventional method; for the latter dataset, our classifier leads to a correct classification rate of 99.67%, 99.67%, and 100.00% for DME, AMD, and normal images, respectively.

이 논문은 고차원의 데이터를 저 차원으로 줄이는 방법 중 하나인 특징추출에 대한 방법들의 특성을 비교한다. 비교대상 방법은 전통적인 PCA(Principal Component Analysis)방법과 시각피질의 특성을 보인다고 알려진 ICA(Independent Component Analysis), 국소기반인식을 구현한 NMF(Non-negative Matrix Factorization), 그리고 이의 성능을 개선한 sNMF(Sparse NMF)로 정하였다. 추출된 특징들의 특성을 시각적으로 확인하기 위하여 필기체 숫자 영상을 대상으로 특징추출을 수행하였으며, 인식기에 적용한 효과의 확인을 위하여 추출된 특징을 다층퍼셉트론에 학습시켜보았다. 각 방법의 특성을 비교한 결과는 응용하고자 하는 문제에서 어떤 특징을 추출하기 원하느냐에 따라 특징추출 방법을 선정할 때 유용할 것이다. In this paper, feature extraction methods, which is one field of reducing dimensions of high-dimensional data, are empirically investigated. We selected the traditional PCA(Principal Component Analysis), ICA(Independent Component Analysis), NMF(Non-negative Matrix Factorization), and sNMF(Sparse NMF) for comparisons. ICA has a similar feature with the simple cell of V1. NMF implemented a "parts-based representation in the brain" and sNMF is a improved version of NMF. In order to visually investigate the extracted features, handwritten digits are handled. Also, the extracted features are used to train multi-layer perceptrons for recognition test. The characteristic of each feature extraction method will be useful when applying feature extraction methods to many real-world problems.

Wavelet-based spatiotemporal sparse quaternion dictionary learning for reconstruction of multi-channel vibration data

Speech can be expressed as a mechanism of expressing thoughts and ideas using vocal sounds. In humans, speech signals are generated at vocal cords, travelled through the vocal tract, and finally produced & transmitted through speaker’s mouth. These speech signals then usually travels through air or other mediums to the listener’s ear, where it acts as pressure waves. The bandwidth of speech signal is around 4 KHz. The noise produced by various ambient sources such as vehicles normally lies in this frequency range. Therefore, speech signals get easily distorted by the ambient noise. These distorted or degraded speech signals are called noisy speech signals. This paper focuses on speech processing (in particularly speech enhancement) of the noisy speech signals. This is very important as speech is the most commonly used way of communication and interaction between humans; however, it is very complex to understand. Therefore, this paper proposes an adaptive algorithm based on Wiener filter for speech enhancement. The proposed adaptive Wiener filter depends on the adaptation of the filter transfer function from sample to sample based on the speech signal statistics (mean and variance). The adaptive Wiener filter is implemented in time domain rather than in frequency domain. This is done to accommodate the random speech signal. The proposed method is compared to the traditional Wiener filter and the spectral subtraction methods.

NMF (non-negative matrix factorization) is a recently addressed speech signal processing method, In this paper, we proposed a new NMF algorithm based on improved PSO (particle swarm optimization) techniques at aims to extract non-negative components with low cross-talking error and high SNR. Compared with standard PSO algorithm, the improved PSO can overcome lower velocity of convergence by updating dynamic inertia weight. Our discussion is supported by experimental results for separating speech signals, which show that the proposed approach exhibits good performance than traditional NMF methods.

Nonnegative matrix factorization (NMF) is known as a parts-based linear representation for nonnegative data. This method has been applied for blind source separation (BSS) when the sources are nonnegative. This paper presents a new NMF method for independent component analysis (ICA), which is useful for BSS without the nonnegativity constraint. Using this method, we transform the sources by their cumulative distribution functions (CDFs) and perform the nonparametric quantization to construct a nonnegative matrix where each entry represents the joint probability density of two transformed signals. The NMF procedure is accordingly realized to find the ICA demixing matrix. The independence between sources is maximized towards attaining the uniformity in the joint probability density. In the experiments on the separation of signal and music signals, we show the effectiveness of the proposed NMF-ICA compared to the infomax ICA and FastICA algorithms.

This paper deals with the application of spatially localized, nonoverlapping features for face recognition. The analysis is carried out by using the features generated from two closely related techniques known as independent component analysis (ICA) and nonnegative matrix factorization (NMF). A set of statistically independent basis vectors with sparse features is derived from ICA. Likewise, NMF is used to yield sparse representation of localized features to represent distributed parts over a human face. Similarities between reconstructed faces of test images and a set of synthesised face representations from the basis vectors derived from an image database using the two techniques are measured. The strengths and weaknesses of each method in the context of face recognition are discussed.

Speech signal is the primary mode of communication amongst humans. It is also used for the communication between human and machine, i.e., human-machine interaction (IoTs), communication with a range of electronic devices, e.g., Internet of Things (IoTs), and even with virtual assistants (VAs) or online search engines, e.g., Google etc. The use of speech signals for the communication over mobile phones, and smart phones based virtual assistants such as Cortana, Siri, Alexa etc. is well known. Usually, this communication involves some communication medium, which is susceptible to noise. The speech or audio signal used for the communication may be adversely affected by the noise that is added unwantedly over the communication channel or medium. It not only degrades the quality of communication and intelligibility of the message or content, but also leads to poor experience. The quality of speech signal may be degraded significantly, due to changes in the characteristics of speech signal, during the process from transmission to reception and the signal conversions involved at both ends. Hence there is a need for algorithms, signal processing methods, acoustic features and classifiers etc. for noise detection and noise reduction during the speech signal communication process. In this paper, a comprehensive review of various speech signal databases, speech signal processing methods used, acoustic features extracted and the classifiers used in different studies is presented. A comprehensive and comparative view of results obtained in different studies, and the applications is also presented. This study should be helpful to all the researchers working in the domains of speech signal processing in noisy environments or degraded speech etc.

Recent studies have shown that brain lesions following stroke can be probabilistically mapped onto disconnections of white matter tracts, and that the resulting “disconnectome” is predictive of the patient’s behavioral deficits. Disconnectome maps are sparse, high-dimensional 3D matrices that require unsupervised dimensionality reduction followed by supervised learning for prediction of the associated behavioral data. However, the optimal machine learning pipeline for disconnectome data still needs to be identified. We examined four dimensionality reduction methods at varying levels of compression and used the extracted features as input for cross-validated regularized regression to predict the associated language and motor deficits. Features extracted by Principal Component Analysis and Non-Negative Matrix Factorization were found to be the best predictors, followed by Independent Component Analysis and Dictionary Learning. Optimizing the number of extracted features improved predictive accuracy and greatly reduced model complexity. Moreover, the choice of dimensionality reduction technique was found to optimally combine with a specific type of regularized regression (ridge vs. LASSO). Overall, our findings represent an important step towards an optimal pipeline that yields high prediction accuracy with a small number of features, which can also improve model interpretability.