Orientadores: João Marcos Travassos Romano, Romis Ribeiro de Faissol Attux / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Eletrica e de Computação / Made available in DSpace on 2018-08-12T15:14:54Z (GMT). No. of bitstreams: 1 Dias_TiagoMacedo_M.pdf: 3219855 bytes, checksum: 5572e53d65cb457f420e78b3150dd6ee (MD5) Previous issue date: 2008 / Resumo: Esta dissertação se propõe a estudar um novo método para separação cega de fontes baseado no modelo Post-Nonlinear, que une uma ferramenta de busca global baseada em computação bioinspirada a uma etapa de busca local conduzida pelo algoritmo FastICA. A idéia subjacente à proposta é procurar obter soluções precisas e eficientes usando de maneira parcimoniosa os recursos computacionais disponíveis. A nova proposta foi testada em diferentes cenários, e, em todos os casos, estabeleceram-se comparações com uma abordagem alternativa, cujo passo de otimização não inclui o estágio de busca local (ou "memética"). Os resultados obtidos por meio de simulações indicam que um bom compromisso entre desempenho e custo computacional foi, de fato, atingido. / Resumo: Esta dissertação se propõe a estudar um novo método para separação cega de fontes baseado no modelo Post-Nonlinear, que une uma ferramenta de busca global baseada em computação bioinspirada a uma etapa de busca local conduzida pelo algoritmo FastICA. A idéia subjacente à proposta é procurar obter soluções precisas e eficientes usando de maneira parcimoniosa os recursos computacionais disponíveis. A nova proposta foi testada em diferentes cenários, e, em todos os casos, estabeleceram-se comparações com uma abordagem alternativa, cujo passo de otimização não inclui o estágio de busca local (ou "memética"). Os resultados obtidos por meio de simulações indicam que um bom compromisso entre desempenho e custo computacional foi, de fato, atingido. / Abstract: This work deals with a new method for source separation of Post-Nonlinear mixtures that brings together an evolutionary-based global search and a local search step based on the FastICA algorithm. The rationale of the proposal is to attempt to obtain efficient and precise solutions using with parsimony the available computational resources. The new proposal was tested in different scenarios and, in all cases, we attempted to establish grounds for comparison with an alternative approach whose optimization step does not include the local (memetic) search stage. Simulation results indicate that a good tradeoff between performance and computational cost was indeed reached. / Abstract: This work deals with a new method for source separation of Post-Nonlinear mixtures that brings together an evolutionary-based global search and a local search step based on the FastICA algorithm. The rationale of the proposal is to attempt to obtain efficient and precise solutions using with parsimony the available computational resources. The new proposal was tested in different scenarios and, in all cases, we attempted to establish grounds for comparison with an alternative approach whose optimization step does not include the local (memetic) search stage. Simulation results indicate that a good tradeoff between performance and computational cost was indeed reached. / Mestrado / Telecomunicações e Telemática / Mestre em Engenharia Elétrica

published_or_final_version / abstract / Electrical and Electronic Engineering / Doctoral / Doctor of Philosophy

Syftet med min undersökning var att ta reda på hur pedagoger tillsammans med barnen på förskolor med Grön Flagg certifikat arbetar med nedskräpningsfrågor och källsortering. Jag använde mig av semistrukturerade intervjuer och sex pedagoger som arbetar på Grön Flagg förskolor. Av resultaten framgår att man på Grön Flagg förskolor arbetar relativt kontinuerligt med nedskräpningsfrågor och källsortering tillsammans med barnen. Pedagogerna anser att det är två bra sätt att arbeta med miljö då det är konkret och något som kan väcka intresse för miljöfrågor hos barnen. Resultatet visar också vikten av att låta barnen vara delaktiga och ha inflytande över arbetet med nedskräpningsfrågor och källsortering. Något som också framkommer av undersökningen är att barnen har blivit mer intresserade och medvetna om miljön tack vare deras arbete med miljöcertifieringen Grön Flagg. / The purpose of this study was to examine how pre-school teachers at pre-schools with Green Flag certification work together with the children concerning questions about littering issues and recycling. I used semi-structured interviews and six teachers were interviewed, all working at pre-schools with Green Flag certificate. The results show that they work relatively constantly with littering issues and source separation with the children. Pre-school teachers believe that these are two good activities because they are tangible and something that can spark interest in environmental issues among children. The results also show the importance of allowing children to be involved and have a say in the work of littering issues and recycling. What also emerges from the survey is that children have become more interested and aware of the environment through their work with Green Flag.

This thesis concerns blind source separation techniques using second order statistics and higher order statistics for reverberant environments. A focus of the thesis is algorithmic simplicity with a view to the algorithms being implemented in their online forms. The main challenge of blind source separation applications is to handle reverberant acoustic environments; a further complication is changes in the acoustic environment such as when human speakers physically move. A novel time-domain method which utilises a pair of finite impulse response filters is proposed. The method of principle angles is defined which exploits a singular value decomposition for their design. The pair of filters are implemented within a generalised sidelobe canceller structure, thus the method can be considered as a beamforming method which cancels one source. An adaptive filtering stage is then employed to recover the remaining source, by exploiting the output of the beamforming stage as a noise reference. A common approach to blind source separation is to use methods that use higher order statistics such as independent component analysis. When dealing with realistic convolutive audio and speech mixtures, processing in the frequency domain at each frequency bin is required. As a result this introduces the permutation problem, inherent in independent component analysis, across the frequency bins. Independent vector analysis directly addresses this issue by modeling the dependencies between frequency bins, namely making use of a source vector prior. An alternative source prior for real-time (online) natural gradient independent vector analysis is proposed. A Student's t probability density function is known to be more suited for speech sources, due to its heavier tails, and is incorporated into a real-time version of natural gradient independent vector analysis. The final algorithm is realised as a real-time embedded application on a floating point Texas Instruments digital signal processor platform. Moving sources, along with reverberant environments, cause significant problems in realistic source separation systems as mixing filters become time variant. A method which employs the pair of cancellation filters, is proposed to cancel one source coupled with an online natural gradient independent vector analysis technique to improve average separation performance in the context of step-wise moving sources. This addresses `dips' in performance when sources move. Results show the average convergence time of the performance parameters is improved. Online methods introduced in thesis are tested using impulse responses measured in reverberant environments, demonstrating their robustness and are shown to perform better than established methods in a variety of situations.

On se place dans le cadre du traitement des signaux audio multicanaux et multi-sources. À partir du mélange de plusieurs sources sonores enregistrées en milieu réverbérant, on cherche à estimer les réponses acoustiques (ou filtres de mélange) entre les sources et les microphones. Ce problème inverse ne peut être résolu qu'en prenant en compte des hypothèses sur la nature des filtres. Notre approche consiste d'une part à identifier mathématiquement les hypothèses nécessaires sur les filtres pour pouvoir les estimer et d'autre part à construire des fonctions de coût et des algorithmes permettant de les estimer effectivement. Premièrement, nous avons considéré le cas où les signaux sources sont connus. Nous avons développé une méthode d'estimation des filtres basée sur une régularisation convexe prenant en compte à la fois la nature parcimonieuse des filtres et leur enveloppe de forme exponentielle décroissante. Nous avons effectué des enregistrements en environnement réel qui ont confirmé l'efficacité de cet algorithme. Deuxièmement, nous avons considéré le cas où les signaux sources sont inconnus, mais statistiquement indépendants. Les filtres de mélange peuvent alors être estimés à une indétermination de permutation et de gain près à chaque fréquence par des techniques d'analyse en composantes indépendantes. Nous avons apporté une étude exhaustive des garanties théoriques par lesquelles l'indétermination de permutation peut être levée dans le cas où les filtres sont parcimonieux dans le domaine temporel. Troisièmement, nous avons commencé à analyser les hypothèses sous lesquelles notre algorithme d'estimation des filtres pourrait être étendu à l'estimation conjointe des signaux sources et des filtres et montré un premier résultat négatif inattendu : dans le cadre de la déconvolution parcimonieuse aveugle, pour une famille assez large de fonctions de coût régularisées, le minimum global est trivial. Des contraintes supplémentaires sur les signaux sources ou les filtres sont donc nécessaires. / This work is focused on the processing of multichannel and multisource audio signals. From an audio mixture of several audio sources recorded in a reverberant room, we wish to estimate the acoustic responses (a.k.a. mixing filters) between the sources and the microphones. To solve this inverse problem one need to take into account additional hypotheses on the nature of the acoustic responses. Our approach consists in first identifying mathematically the necessary hypotheses on the acoustic responses for their estimation and then building cost functions and algorithms to effectively estimate them. First, we considered the case where the source signals are known. We developed a method to estimate the acoustic responses based on a convex regularization which exploits both the temporal sparsity of the filters and the exponentially decaying envelope. Real-world experiments confirmed the effectiveness of this method on real data. Then, we considered the case where the sources signal are unknown, but statistically independent. The mixing filters can be estimated up to a permutation and scaling ambiguity. We brought up an exhaustive study of the theoretical conditions under which we can solve the indeterminacy, when the multichannel filters are sparse in the temporal domain. Finally, we started to analyse the hypotheses under which this algorithm could be extended to the joint estimation of the sources and the filters, and showed a first unexpected results : in the context of blind deconvolution with sparse priors, for a quite large family of regularised cost functions, the global minimum is trivial. Additional constraints on the source signals and the filters are needed.

Thanks to its ability to evaluate metabolic functions in tissues from the temporal evolution of a previously injected radiotracer, dynamic positron emission tomography (PET) has become an ubiquitous analysis tool to quantify biological processes. Several quantification techniques from the PET imaging literature require a previous estimation of global time-activity curves (TACs) (herein called \textit{factors}) representing the concentration of tracer in a reference tissue or blood over time. To this end, factor analysis has often appeared as an unsupervised learning solution for the extraction of factors and their respective fractions in each voxel. Inspired by the hyperspectral unmixing literature, this manuscript addresses two main drawbacks of general factor analysis techniques applied to dynamic PET. The first one is the assumption that the elementary response of each tissue to tracer distribution is spatially homogeneous. Even though this homogeneity assumption has proven its effectiveness in several factor analysis studies, it may not always provide a sufficient description of the underlying data, in particular when abnormalities are present. To tackle this limitation, the models herein proposed introduce an additional degree of freedom to the factors related to specific binding. To this end, a spatially-variant perturbation affects a nominal and common TAC representative of the high-uptake tissue. This variation is spatially indexed and constrained with a dictionary that is either previously learned or explicitly modelled with convolutional nonlinearities affecting non-specific binding tissues. The second drawback is related to the noise distribution in PET images. Even though the positron decay process can be described by a Poisson distribution, the actual noise in reconstructed PET images is not expected to be simply described by Poisson or Gaussian distributions. Therefore, we propose to consider a popular and quite general loss function, called the $\beta$-divergence, that is able to generalize conventional loss functions such as the least-square distance, Kullback-Leibler and Itakura-Saito divergences, respectively corresponding to Gaussian, Poisson and Gamma distributions. This loss function is applied to three factor analysis models in order to evaluate its impact on dynamic PET images with different reconstruction characteristics.

Neste trabalho, investigam-se técnicas autodidatas baseadas em estatísticas de ordem superior, aplicadas à equalização de canais de comunicação. Inicialmente, obtém-se um intervalo do passo de adaptação que assegura a convergência do algoritmo do Módulo Constante com o gradiente exato. Algoritmos como o CMA (Constant Modulus Algorithm) e o SWA (Shalvi-Weinstein Algorithm) são revisitados e suas capacidades de tracking analisadas, utilizando-se uma relação de conservação de energia. Além disso, é proposto um algoritmo autodidata denominado AC-CMA (Accelerated Constant Modulus Algorithm) que utiliza a segunda derivada (“aceleração") da estimativa dos coeficientes. Esse algoritmo pode apresentar um compromisso mais favorável entre complexidade computacional e velocidade de convergência que o CMA e o SWA. Esses resultados são estendidos para o caso multiusuário. Através de simulações, os algoritmos são comparados e as análises de convergência e tracking validadas. Considerando o DFE (Decision Feedback Equalizer) no caso monousuário com o critério do módulo constante, é proposto um algoritmo concorrente que evita soluções degeneradas e apresenta um desempenho melhor do que os existentes na literatura. Com o intuito de evitar propagação de erros, é proposta uma estrutura híbrida que utiliza uma rede neural recorrente na malha de realimentação. Resultados de simulações indicam que seu uso pode ser vantajoso para canais lineares e não-lineares. / The equalization of communication channels is addressed by using blind techniques based on higher order statistics. A step-size interval is obtained to ensure the convergence of Steepest-Descent Constant Modulus Algorithm. The Shalvi-Weinstein Algorithm (SWA) and Constant Modulus Algorithm (CMA) are revisited and their tracking capabilities are analyzed by using an energy conservation relation. Moreover, a novel blind algorithm named Accelerated Constant Modulus Algorithm (AC-CMA) is proposed. It adjusts the second derivative (“acceleration") of the coefficient estimates and presents a more favorable compromise between computational complexity and convergence rate than CMA or SWA. These results are extended to the MIMO (Multiple-Input Multiple-Output) case. By means of simulations, the algorithms are compared and the convergence and tracking analysis are validated. The Decision Feedback Equalizer (DFE) is considered in the SISO (Single-Input Single-Output) case with the Constant Modulus criterion and a concurrent algorithm is proposed. It avoids degenerated solutions and shows better behavior than the others presented in the literature. In order to avoid error propagation, a hybrid DFE is also proposed. It includes a recurrent neural network in the feedback filter and may be advantageously used to equalize linear and nonlinear channels.

Although distinguishing different sounds in noisy environment is a relative easy task for human, source separation has long been extremely difficult in audio signal processing. The problem is challenging for three reasons: the large variety of sound type, the abundant mixing conditions and the unclear mechanism to distinguish sources, especially for similar sounds. In recent years, the neural network based methods achieved impressive successes in various problems, including the speech enhancement, where the task is to separate the clean speech out of the noise mixture. However, the current deep learning based source separator does not perform well on real recorded noisy speech, and more importantly, is not applicable in a more general source separation scenario such as overlapped speech. In this thesis, we firstly propose extensions for the current mask learning network, for the problem of speech enhancement, to fix the scale mismatch problem which is usually occurred in real recording audio. We solve this problem by combining two additional restoration layers in the existing mask learning network. We also proposed a residual learning architecture for the speech enhancement, further improving the network generalization under different recording conditions. We evaluate the proposed speech enhancement models on CHiME 3 data. Without retraining the acoustic model, the best bi-direction LSTM with residue connections yields 25.13% relative WER reduction on real data and 34.03% WER on simulated data. Then we propose a novel neural network based model called “deep clustering” for more general source separation tasks. We train a deep network to assign contrastive embedding vectors to each time-frequency region of the spectrogram in order to implicitly predict the segmentation labels of the target spectrogram from the input mixtures. This yields a deep network-based analogue to spectral clustering, in that the embeddings form a low-rank pairwise affinity matrix that approximates the ideal affinity matrix, while enabling much faster performance. At test time, the clustering step “decodes” the segmentation implicit in the embeddings by optimizing K-means with respect to the unknown assignments. Experiments on single channel mixtures from multiple speakers show that a speaker-independent model trained on two-speaker and three speakers mixtures can improve signal quality for mixtures of held-out speakers by an average over 10dB. We then propose an extension for deep clustering named “deep attractor” network that allows the system to perform efficient end-to-end training. In the proposed model, attractor points for each source are firstly created the acoustic signals which pull together the time-frequency bins corresponding to each source by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. We showed that this frame work can achieve even better results. Lastly, we introduce two applications of the proposed models, in singing voice separation and the smart hearing aid device. For the former, a multi-task architecture is proposed, which combines the deep clustering and the classification based network. And a new state of the art separation result was achieved, where the signal to noise ratio was improved by 11.1dB on music and 7.9dB on singing voice. In the application of smart hearing aid device, we combine the neural decoding with the separation network. The system firstly decodes the user’s attention, which is further used to guide the separator for the targeting source. Both objective study and subjective study show the proposed system can accurately decode the attention and significantly improve the user experience.

Neste trabalho, investigam-se técnicas autodidatas baseadas em estatísticas de ordem superior, aplicadas à equalização de canais de comunicação. Inicialmente, obtém-se um intervalo do passo de adaptação que assegura a convergência do algoritmo do Módulo Constante com o gradiente exato. Algoritmos como o CMA (Constant Modulus Algorithm) e o SWA (Shalvi-Weinstein Algorithm) são revisitados e suas capacidades de tracking analisadas, utilizando-se uma relação de conservação de energia. Além disso, é proposto um algoritmo autodidata denominado AC-CMA (Accelerated Constant Modulus Algorithm) que utiliza a segunda derivada (“aceleração”) da estimativa dos coeficientes. Esse algoritmo pode apresentar um compromisso mais favorável entre complexidade computacional e velocidade de convergência que o CMA e o SWA. Esses resultados são estendidos para o caso multiusuário. Através de simulações, os algoritmos são comparados e as análises de convergência e tracking validadas. Considerando o DFE (Decision Feedback Equalizer) no caso monousuário com o critério do módulo constante, é proposto um algoritmo concorrente que evita soluções degeneradas e apresenta um desempenho melhor do que os existentes na literatura. Com o intuito de evitar propagação de erros, é proposta uma estrutura híbrida que utiliza uma rede neural recorrente na malha de realimentação. Resultados de simulações indicam que seu uso pode ser vantajoso para canais lineares e não-lineares. / The equalization of communication channels is addressed by using blind techniques based on higher order statistics. A step-size interval is obtained to ensure the convergence of Steepest-Descent Constant Modulus Algorithm. The Shalvi-Weinstein Algorithm (SWA) and Constant Modulus Algorithm (CMA) are revisited and their tracking capabilities are analyzed by using an energy conservation relation. Moreover, a novel blind algorithm named Accelerated Constant Modulus Algorithm (AC-CMA) is proposed. It adjusts the second derivative (“acceleration”) of the coefficient estimates and presents a more favorable compromise between computational complexity and convergence rate than CMA or SWA. These results are extended to the MIMO (Multiple-Input Multiple-Output) case. By means of simulations, the algorithms are compared and the convergence and tracking analysis are validated. The Decision Feedback Equalizer (DFE) is considered in the SISO (Single-Input Single-Output) case with the Constant Modulus criterion and a concurrent algorithm is proposed. It avoids degenerated solutions and shows better behavior than the others presented in the literature. In order to avoid error propagation, a hybrid DFE is also proposed. It includes a recurrent neural network in the feedback filter and may be advantageously used to equalize linear and nonlinear channels.

Automatic speech recognition has become a standard feature on many consumer electronics and automotive products, and the accuracy of the decoded speech has improved dramatically over time. Often, designers of these products achieve accuracy by employing microphone arrays and beamforming algorithms to reduce interference. However, beamforming microphone arrays are too large for small form factor products such as smart watches. Yet these small form factor products, which have precious little space for tactile user input (i.e. knobs, buttons and touch screens), would benefit immensely from a user interface based on reliably accurate automatic speech recognition. This thesis proposes a solution for interference mitigation that employs blind source separation with a compact array of commercially available unidirectional microphone elements. Such an array provides adequate spatial diversity to enable blind source separation and would easily fit in a smart watch or similar small form factor product. The solution is characterized using publicly available speech audio clips recorded for the purpose of testing automatic speech recognition algorithms. The proposal is modelled in different interference environments and the efficacy of the solution is evaluated. Factors affecting the performance of the solution are identified and their influence quantified. An expectation is presented for the quality of separation as well as the resulting improvement in word error rate that can be achieved from decoding the separated speech estimate versus the mixture obtained from a single unidirectional microphone element. Finally, directions for future work are proposed, which have the potential to improve the performance of the solution thereby making it a commercially viable product.