This research is mainly concerned with a robust method for improving the performance of a real-time speech enhancement and noise cancellation for Automatic Speech Recognition (ASR) in a real-time environment. Therefore, the thesis titled, “Real-time adaptive beamformer for Automatic speech Recognition in a car environment” presents an application technique of a beamforming method and Automatic Speech Recognition (ASR) method. In this thesis, a novel solution is presented to the question as below, namely: How can the driver’s voice control the car using ASR? The solution in this thesis is an ASR using a hybrid system with acoustic beamforming Voice Activity Detector (VAD) and an Adaptive Wiener Filter. The beamforming approach is based on a fundamental theory of normalized least-mean squares (NLMS) to improve Signal to Noise Ratio (SNR). The microphone has been implemented with a Voice Activity Detector (VAD) which uses time-delay estimation together with magnitude-squared coherence (MSC). An experiment clearly shows the ability of the composite system to reduce noise outside of a defined active zone. In real-time environments a speech recognition system in a car has to receive the driver’s voice only whilst suppressing background noise e.g. voice from radio. Therefore, this research presents a hybrid real-time adaptive filter which operates within a geometrical zone defined around the head of the desired speaker. Any sound outside of this zone is considered to be noise and suppressed. As this defined geometrical zone is small, it is assumed that only driver's speech is incoming from this zone. The technique uses three microphones to define a geometric based voice-activity detector (VAD) to cancel the unwanted speech coming from outside of the zone. In the case of a sole unwanted speech incoming from outside of a desired zone, this speech is muted at the output of the hybrid noise canceller. In case of an unwanted speech and a desired speech are incoming at the same time, the proposed VAD fails to identify the unwanted speech or desired speech. In such a situation an adaptive Wiener filter is switched on for noise reduction, where the SNR is improved by as much as 28dB. In order to identify the signal quality of the filtered signal from Wiener filter, a template matching speech recognition system that uses a Wiener filter is designed for testing. In this thesis, a commercial speech recognition system is also applied to test the proposed beamforming based noise cancellation and the adaptive Wiener filter.

Orientador: Eurípedes Guilherme de Oliveira Nóbrega / Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Mecânica / Made available in DSpace on 2018-08-16T17:08:16Z (GMT). No. of bitstreams: 1 AlvesJunior_MarcoAntoniodeOliveira_D.pdf: 2426154 bytes, checksum: bc1d1ecd78ca881b519a4e56c9383c8f (MD5) Previous issue date: 2010 / Resumo: Esta monografia apresenta uma arquitetura para aplicação de Controle Tolerante a Falhas, seguindo uma abordagem de dupla malha de realimentação. A primeira malha apresenta um controlador regular, e a segunda segue uma estratégia de controle adaptativa baseada em rede neural, que faz uso de um mecanismo de ajustes de pesos em tempo real. O primeiro controlador foi escolhido como um projeto de controlador baseado em norma H?, objetivando estabilizar o sistema e garantir o bom desempenho na presença de erros de modelagem e distúrbios externos. O controlador tolerante a falhas, que atua complementarmente à malha externa, é o controlador que usa a técnica neuroadaptativa. A rede neural possui estados internos recorrentes, usando uma superfície de deslizamento para adaptar os seus pesos, de modo a acomodar as possíveis falhas. Também apresenta robustez contra as perturbações externas, além da capacidade do controlador regular. Uma nova topologia de dinâmica da rede neural, com estados internos recursivos e aprendizado em tempo-real, é proposta, e a estabilidade do sistema é provada com base em uma função de Lyapunov e em requisitos predefinidos. Para avaliar o método, foi usado um modelo matemático de um veículo aéreo não tripulado do tipo quadrirrotor. Os resultados simulados, com o sistema submetido a vários tipos de condições de falha, são apresentados, mostrando o bom desempenho da configuração proposta / Abstract: This monograph presents an architecture scheme for Fault Tolerant Control applications, following a dual-loop controller design approach, where the first closed loop is a regular controller and the second one is based on a neural network adaptive control strategy, with on-line adjustment of the weights. The first controller, which was here chosen as an H? norm designed controller, aims stabilize the system, and guarantee a good performance in presence of modeling errors and external disturbances. The fault tolerant controller, acting complementarily to the external loop, is the one using the neuro-adaptive technique. Its design is based on recurrent internal states, using a sliding surface to adapt the weights of the neural network, in order to accommodate the system faults, but also with a robust effect which includes correcting all external perturbations, beyond the capacity of the regular controller. A new neural network dynamic topology, with internal recursive states and on-line learning algorithm, is proposed, and its stability is proved based on a Lyapunov function and predefined requirements. To assess the method, an unmanned quad rotor flying vehicle is modeled, and the respective controllers designed. Results based on numerical simulation, with the system submitted to several different fault conditions, are presented, showing a good performance of the proposed configuration / Doutorado / Mecanica dos Sólidos e Projeto Mecanico / Doutor em Engenharia Mecânica

Orientadores: João Marcos Travassos Romano, Hisham Abou-Kandil, Mohamed Abbas-Turki / Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia Elétrica e de Computação e ENS-Cachan (École Normale Supérieure de Cachan) / Made available in DSpace on 2018-08-18T08:26:12Z (GMT). No. of bitstreams: 1 Chaves_FabianodeSousa_D.pdf: 2605316 bytes, checksum: 88049a9675c3cd7f4955cc1925e8bd01 (MD5) Previous issue date: 2010 / Resumo: A interferência é um dos fatores limitantes do desempenho individual e global em redes de comunicação sem fio. Neste trabalho, duas técnicas clássicas de gerenciamento de interferência são estudadas: o controle de potência de transmissão e a equalização de canal. Três abordagens são consideradas para o controle de potência distribuído e oportunista. A primeira tem por base a teoria dos jogos estáticos não-cooperativos e teorias de funções iterativas, resultando em uma classe de algoritmos. Na segunda abordagem, propomos diferentes algoritmos derivados de formulações e soluções tradicionais dos controles H2 e misto H2/Hoo.Por fim, aplicamos a teoria dos jogos dinâmicos ao problema para a obtenção de dois novos algoritmos de controle de potência. A segunda parte da tese, que trata do problema da equalização, é dividida em dois tópicos. No primeiro, fazemos uma análise de "pior caso" do atraso de equalização por meio de conceitos da teoria dos jogos não-cooperativos. No segundo tópico, apresentamos duas propostas para a reunião das características desejáveis dos equalizadores H2 e Hoo: uma combinação convexa dos dois filtros e um esquema de adaptação do nível de robustez do filtro Hoo. / Abstract: Interference is a limiting factor of individual and global performance in wireless communication networks. In this work, two classical interference management techniques are studied: the transmission power control and the channel equalization. Three approaches are considered for distributed and opportunistic power control. The first one is based on static non-cooperative game theory and theories of iterative functions, providing a class of algorithms. In the second approach, we propose different algorithms derived from formulations and traditional solutions of H2 control and mixed H2/Hoo control. Finally, we apply dynamic game theory to the problem for obtaining two new power control algorithms. The second part of the thesis, devoted to channel equalization, is divided into two topics. In the first one, we provide a "worst case" analysis for equalization delay by using concepts of noncooperative game theory. In the second topic, we present two proposals for the combination of the desirable characteristics of H2 and Hoo equalizers: a convex combination of the two filters and a scheme for adapting the robustness level of the Hoo filter. / Doutorado / Telecomunicações e Telemática / Doutor em Engenharia Elétrica

Made available in DSpace on 2016-08-17T14:53:26Z (GMT). No. of bitstreams: 1 Dissertacao Aleksandro Costa.pdf: 1706153 bytes, checksum: 8d61027896dbab484303f78ed17b9b70 (MD5) Previous issue date: 2014-02-28 / FUNDAÇÃO DE AMPARO À PESQUISA E AO DESENVOLVIMENTO CIENTIFICO E TECNOLÓGICO DO MARANHÃO / This work has the objective to develop an analytical model that makes prediction of the behavior of the algorithm as a function of the design parameters (step adaptation, kernel function and its parameters).We use a non-quadratic function based on kernel, performing a nonlinear transformation of the input space filtering applied on line. Was developed and implemented in the system for adaptive filtering based on Kernel, which provides an analysis of the behavior of KRLS algorithm as well as its properties of convergence. It applies a kernel function in the cost function from the non-recursive quadratic function of an even power, which minimizes the error, defined as the expectation of the cumulative cost of actions taken along a sequence of steps. It appears that this approach allows the determination of the parameters of the problem with greater reliability and robustness and lower cost compared with traditional algorithms (RLS, KRLS, RNQ) . / Este trabalho tem como objetivo desenvolver um modelo analítico que faça a previsão do comportamento do algoritmo RLS como uma função dos parâmetros de projeto (passo de adaptação, função kernel e seus parâmetros). Utiliza-se uma função não quadrática baseado em kernel, realizando uma transformação não linear do espaço de entrada aplicada à filtragem. Foi desenvolvido e implementado na redução de ruídos para a filtragem adaptativa baseada em Kernel, que fornece uma análise do comportamento do algoritmo KRLS, bem como de suas propriedades de convergência. Aplica-se uma função kernel na função de custo a partir da função recursiva não quadrática de quarta potência, que minimiza o erro, definido como a expectativa do custo cumulativo de ações tomadas ao longo de uma sequência de passos. Verifica-se que essa abordagem possibilita a determinação dos parâmetros do problema com uma maior confiabilidade e robustez e o menor custo, quando comparado com algoritmos tradicionais (RLS, KRLS, RNQ).

Submitted by PPG Engenharia El?trica (engenharia.pg.eletrica@pucrs.br) on 2018-10-29T13:30:23Z No. of bitstreams: 1 TIAGO BENETTI_DIS.pdf: 5038519 bytes, checksum: 95fa8d1b367b574eee27e772a55a9a49 (MD5) / Approved for entry into archive by Caroline Xavier (caroline.xavier@pucrs.br) on 2018-10-30T17:21:55Z (GMT) No. of bitstreams: 1 TIAGO BENETTI_DIS.pdf: 5038519 bytes, checksum: 95fa8d1b367b574eee27e772a55a9a49 (MD5) / Made available in DSpace on 2018-10-30T17:27:25Z (GMT). No. of bitstreams: 1 TIAGO BENETTI_DIS.pdf: 5038519 bytes, checksum: 95fa8d1b367b574eee27e772a55a9a49 (MD5) Previous issue date: 2018-08-31 / Heart rate monitoring using Photoplethysmography (PPG) signals acquired from the individuals pulse has become popular due to emergence of numerous low cost wearable devices. However, monitoring during physical activities has obstacles because of the influence of motion artifacts in PPG signals. The objective of this work is to introduce a new algorithm capable of removing motion artifacts and estimating heart rate from pulse PPG signals. Normalized Least Mean Square (NLMS) and Recursive Least Squares (RLS) algorithms are proposed for an adaptive filtering structure that uses acceleration signals as reference to remove motion artifacts. The algorithm uses the Periodogram of the filtered signals to extract their heart rates, which will be used together with a PPG Signal Quality Index to feed the input of a Kalman Filter. Specific heuristics and the Quality Index collaborate so that the Kalman filter provides a heart rate estimate with high accuracy and robustness to measurement uncertainties. The algorithm was validated from the heart rate obtained from Electrocardiography signals and the proposed method with the RLS algorithm presented the best results with an absolute mean error of 1.54 beats per minute (bpm) and standard deviation of 0.62 bpm, recorded for 12 individuals performing a running activity on a treadmill with varying speeds. The results make the performance of the algorithm comparable and even better than several recently developed methods in this field. In addition, the algorithm presented a low computational cost and suitable to the time interval in which the heart rate estimate is performed. Thus, it is expected that this algorithm will improve the obtaining of heart rate in currently available wearable devices. / O monitoramento da frequ?ncia card?aca utilizando sinais de Fotopletismografia ou PPG (do ingl?s, Photopletismography) adquiridos do pulso de indiv?duos tem se popularizado devido ao surgimento de in?meros dispositivos wearable de baixo custo. No entanto, o monitoramento durante atividades f?sicas tem dificuldades em raz?o da influ?ncia de artefatos de movimento nos sinais de PPG. O objetivo deste trabalho ? introduzir um novo algoritmo capaz de remover artefatos de movimento e estimar a frequ?ncia card?aca de sinais de PPG de pulso. Os algoritmos do M?nimo Quadrado M?dio Normalizado ou NLMS (do ingl?s, Normalized Least Mean Square) e de M?nimos Quadrados Recursivos ou RLS (do ingl?s, Recursive Least Squares) s?o propostos para uma estrutura de filtragem adaptativa que utiliza sinais de acelera??o como refer?ncia para remover os artefatos de movimento. O algoritmo utiliza o Periodograma dos sinais filtrados para extrair suas frequ?ncias card?acas, que ser?o utilizadas juntamente com um ?ndice de Qualidade do Sinal de PPG para alimentar a entrada de um Filtro de Kalman. Heur?sticas espec?ficas e o ?ndice de Qualidade colaboram para que filtro de Kalman forne?a uma estimativa da frequ?ncia card?aca com alta acur?cia e robustez a incertezas de medi??o. O algoritmo foi validado a partir da frequ?ncia card?aca obtida de sinais de Eletrocardiografia e o m?todo proposto com o algoritmo RLS apresentou os melhores resultados com um erro m?dio absoluto de 1,54 batimentos por minuto (bpm) e desvio padr?o de 0,62 bpm, registrados para 12 indiv?duos realizando uma atividade de corrida em uma esteira com velocidades variadas. Os resultados tornam o desempenho do algoritmo compar?vel e at? mesmo melhor que v?rios m?todos desenvolvidos recentemente neste campo. Al?m disso, o algoritmo apresentou um custo computacional baixo e adequado ao intervalo de tempo em que a estimativa da frequ?ncia card?aca ? realizada. Dessa forma, espera-se que este algoritmo melhore a obten??o da frequ?ncia card?aca em dispositivos wearable atualmente dispon?veis.

In system identification scenario, the linear approximation of the system modelled by its impulse response, is estimated in real time by gradient type Least Mean Square (LMS) or Recursive Least Squares (RLS) algorithms. In recent applications like acoustic echo cancellation, the order of the impulse response to be estimated is very high, and these traditional approaches are inefficient and real time implementation becomes difficult. Alternatively, the system is modelled by a set of shorter adaptive filters operating in parallel on subsampled signals. This approach, referred to as subband adaptive filtering, is expected to reduce not only the computational complexity but also to improve the convergence rate of the adaptive algorithm. But in practice, different subband adaptive algorithms have to be used to enhance the performance with respect to complexity, convergence rate and processing delay. A single subband adaptive filtering algorithm which outperforms the full band scheme in all applications is yet to be realized. This thesis is intended to study the subband adaptive filtering techniques and explore the possibilities of better algorithms for performance improvement. Three different subband adaptive algorithms have been proposed and their performance have been verified through simulations. These algorithms have been applied to acoustic echo cancellation and EEG artefact minimization problems. Details of the work To start with, the fast FIR filtering scheme introduced by Mou and Duhamel has been generalized. The Perfect Reconstruction Filter Bank (PRFB) is used to model the linear FIR system. The structure offers efficient implementation with reduced arithmetic complexity. By using a PRFB with non adjacent filters non overlapping, many channel filters can be eliminated from the structure. This helps in reducing the complexity of the structure further, but introduces approximation in the model. The modelling error depends on the stop band attenuation of the filters of the PRFB. The error introduced due to approximation is tolerable for applications like acoustic echo cancellation. The filtered output of the modified generalized fast filtering structure is given by (formula) where, Pk(z) is the main channel output, Pk,, k+1 (z) is the output of auxiliary channel filters at the reduced rate, Gk (z) is the kth synthesis filter and M the number of channels in the PRFB. An adaptation scheme is developed for adapting the main channel filters. Auxiliary channel filters are derived from main channel filters. Secondly, the aliasing problem of the classical structure is reduced without using the cross filters. Aliasing components in the estimated signal results in very poor steady state performance in the classical structure. Attempts to eliminate the aliasing have reduced the computation gain margin and the convergence rate. Any attempt to estimate the subband reference signals from the aliased subband input signals results in aliasing. The analysis filter Hk(z) having the following antialiasing property (formula) can avoid aliasing in the input subband signal. The asymmetry of the frequency response prevents the use of real analysis filters. In the investigation presented in this thesis, complex analysis filters and real'synthesis filters are used in the classical structure, to reduce the aliasing errors and to achieve superior convergence rate. PRFB is traditionally used in implementing Interpolated FIR (IFIR) structure. These filters may not be ideal for processing an input signal for an adaptive algorithm. As third contribution, the IFIR structure is modified using discrete finite frames. The model of an FIR filter s is given by Fc, with c = Hs. The columns of the matrix F forms a frame with rows of H as its dual frame. The matrix elements can be arbitrary except that the transformation should be implementable as a filter bank. This freedom is used to optimize the filter bank, with the knowledge of the input statistics, for initial convergence rate enhancement . Next, the proposed subband adaptive algorithms are applied to acoustic echo cancellation problem with realistic parameters. Speech input and sufficiently long Room Impulse Response (RIR) are used in the simulations. The Echo Return Loss Enhancement (ERLE)and the steady state error spectrum are used as performance measures to compare these algorithms with the full band scheme and other representative subband implementations. Finally, Subband adaptive algorithm is used in minimization of EOG (Electrooculogram) artefacts from measured EEG (Electroencephalogram) signal. An IIR filterbank providing sufficient isolation between the frequency bands is used in the modified IFIR structure and this structure has been employed in the artefact minimization scheme. The estimation error in the high frequency range has been reduced and the output signal to noise ratio has been increased by a couple of dB over that of the fullband scheme. Conclusions Efforts to find elegant Subband adaptive filtering algorithms will continue in the future. However, in this thesis, the generalized filtering algorithm could offer gain in filtering complexity of the order of M/2 and reduced misadjustment . The complex classical scheme offered improved convergence rate, reduced misadjustment and computational gains of the order of M/4 . The modifications of the IFIR structure using discrete finite frames made it possible to eliminate the processing delay and enhance the convergence rate. Typical performance of the complex classical case for speech input in a realistic scenario (8 channel case), offers ERLE of more than 45dB. The subband approach to EOG artefact minimization in EEG signal was found to be superior to their fullband counterpart. (Refer PDF file for Formulas)

No description available.

L’estimation fonctionnelle basée sur les espaces de Hilbert à noyau reproduisant demeure un sujet de recherche actif pour l’identification des systèmes non linéaires. L'ordre du modèle croit avec le nombre de couples entrée-sortie, ce qui rend cette méthode inadéquate pour une identification en ligne. Le critère de cohérence est une méthode de parcimonie pour contrôler l’ordre du modèle. Le modèle est donc défini à partir d'un dictionnaire de faible taille qui est formé par les fonctions noyau les plus pertinentes.Une fonction noyau introduite dans le dictionnaire y demeure même si la non-stationnarité du système rend sa contribution faible dans l'estimation de la sortie courante. Il apparaît alors opportun d'adapter les éléments du dictionnaire pour réduire l'erreur quadratique instantanée et/ou mieux contrôler l'ordre du modèle.La première partie traite le sujet des algorithmes adaptatifs utilisant le critère de cohérence. L'adaptation des éléments du dictionnaire en utilisant une méthode de gradient stochastique est abordée pour deux familles de fonctions noyau. Cette partie a un autre objectif qui est la dérivation des algorithmes adaptatifs utilisant le critère de cohérence pour identifier des modèles à sorties multiples.La deuxième partie introduit d'une manière abrégée le palier magnétique actif (PMA). La proposition de contrôler un PMA par un algorithme adaptatif à noyau est présentée pour remplacer une méthode utilisant les réseaux de neurones à couches multiples / Function approximation methods based on reproducing kernel Hilbert spaces are of great importance in kernel-based regression. However, the order of the model is equal to the number of observations, which makes this method inappropriate for online identification. To overcome this drawback, many sparsification methods have been proposed to control the order of the model. The coherence criterion is one of these sparsification methods. It has been shown possible to select a subset of the most relevant passed input vectors to form a dictionary to identify the model.A kernel function, once introduced into the dictionary, remains unchanged even if the non-stationarity of the system makes it less influent in estimating the output of the model. This observation leads to the idea of adapting the elements of the dictionary to obtain an improved one with an objective to minimize the resulting instantaneous mean square error and/or to control the order of the model.The first part deals with adaptive algorithms using the coherence criterion. The adaptation of the elements of the dictionary using a stochastic gradient method is presented for two types of kernel functions. Another topic is covered in this part which is the implementation of adaptive algorithms using the coherence criterion to identify Multiple-Outputs models.The second part introduces briefly the active magnetic bearing (AMB). A proposed method to control an AMB by an adaptive algorithm using kernel methods is presented to replace an existing method using neural networks

No description available.

In automatic speech recognition systems (ASRs), training is a critical phase to the system?s success. Communication media, either analog (such as analog landline phones) or digital (VoIP) distort the speaker?s speech signal often in very complex ways: linear distortion occurs in all channels, either in the magnitude or phase spectrum. Non-linear but time-invariant distortion will always appear in all real systems. In digital systems we also have network effects which will produce packet losses and delays and repeated packets. Finally, one cannot really assert what path a signal will take, and so having error or distortion in between is almost a certainty. The channel introduces an acoustical mismatch between the speaker's signal and the trained data in the ASR, which results in poor recognition performance. The approach so far, has been to try to undo the havoc produced by the channels, i.e. compensate for the channel's behavior. In this thesis, we try to characterize the effects of different transmission media and use that as an inexpensive and repeatable way to train ASR systems.