Search

We study the collective dynamics of a Leaky Integrate and Fire network in which precise relative phase relationship of spikes among neurons are stored, as attractors of the dynamics, and selectively replayed at differentctime scales. Using an STDP-based learning process, we store in the connectivity several phase-coded spike patterns, and we find that, depending on the excitability of the network, different working regimes are possible, with transient or persistent replay activity induced by a brief signal. We introduce an order parameter to evaluate the similarity between stored and recalled phase-coded pattern, and measure the storage capacity. Modulation of spiking thresholds during replay changes the frequency of the collective oscillation or the number of spikes per cycle, keeping preserved the phases relationship. This allows a coding scheme in which phase, rate and frequency are dissociable. Robustness with respect to noise and heterogeneity of neurons parameters is studied, showing that, since dynamics is a retrieval process, neurons preserve stablecprecise phase relationship among units, keeping a unique frequency of oscillation, even in noisy conditions and with heterogeneity of internal parameters of the units.

In this paper a formerly proposed continuous digit recognition system based on Reservoir Computing (RC) is improved in two respects: (1)the single reservoir is substituted by a stack of reservoirs, and (2)the straightforward mapping of reservoir outputs to state likelihoods is replaced by a trained non-parametric mapping. Furthermore, it is shown that a reservoir-based method can improve a model trained on clean speech to work better in a noisy condition from which it has a number of unknown digit string recordings available. The first two improvements have lead to a system that outperforms a HMM-based system with the same noise robust features as input. The model adaptation offers a promising supplementary gain at modest noise levels.

International audience; A pattern descriptor invariant to rotation, scaling, translation (RST), and robust to additive noise is proposed by using the Radon, Fourier, and Mellin transforms. The Radon transform converts the RST transformations applied on a pattern image into transformations in the radial and angular coordinates of the pattern's Radon image. These beneficial properties of the Radon transform make it an useful intermediate representation for the extraction of invariant features from pattern images for the purpose of indexing/matching. In this paper, invariance to RST is obtained by applying the 1D Fourier-Mellin and discrete Fourier transforms on the radial and angular coordinates of the pattern's Radon image respectively. The implementation of the proposed descriptor is reasonably fast and correct, based mainly on the fusion of the Radon and Fourier transforms and on a modification of the Mellin transform. Theoretical arguments validate the robustness of the proposed descriptor to additive noise and empirical evidence on both occlusion/deformation and noisy datasets shows its effectiveness.

International audience; Path openings and closings are efficient morphological operators that use flexible oriented paths as structuring elements. They are employed in a similar way to operators with rotated line segments as structuring elements, but are more effective at detecting linear structures that are not necessarily locally perfectly straight. While their theory has always allowed paths in arbitrary dimensions, de facto implementations were only proposed in 2D. Recently, a new implementation was proposed enabling the computation of efficient -dimensional path operators. However this implementation is limited in the sense that it is not robust to noise. Indeed, in practical applications, for path operators to be effective, structuring elements must be sufficiently long so that they correspond to the length of the desired features to be detected. Yet, path operators are increasingly sensitive to noise as their length parameter increases. To cope with this limitation, we propose an efficient -dimensional algorithm, the Robust Path Operator, which uses a larger and more flexible family of flexible structuring elements. Given an arbitrary length parameter G, path propagation is allowed if disconnections between two pixels belonging to a path is less or equal to G and so, render it independent of . This simple assumption leads to constant memory bookkeeping and results in a low complexity.

ISBN: 978-1-61284-432-9; International audience; The beneficial properties of the Radon transform make it an useful intermediate representation for the extraction of invariant features from pattern images for the purpose of indexing/matching. This paper revisits the problem with a generic view on a popular Radon-based pattern descriptor, the R-signature, bringing in a class of descriptors spatially describing patterns at all the directions and at different levels. The domain of this class and the selection of its representative are also discussed. Theoretical arguments validate the robustness of the generic R-signature to additive noise and experimental results show its effectiveness.

This paper proposes to use exemplar-based sparse representations for noise robust automatic speech recognition. First, we describe how speech can be modeled as a linear combination of a small number of exemplars from a large speech exemplar dictionary. The exemplars are time-frequency patches of real speech, each spanning multiple time frames. We then propose to model speech corrupted by additive noise as a linear combination of noise and speech exemplars, and we derive an algorithm for recovering this sparse linear combination of exemplars from the observed noisy speech. We describe how the framework can be used for doing hybrid exemplar-based/HMM recognition by using the exemplar-activations together with the phonetic information associated with the exemplars. As an alternative to hybrid recognition, the framework also allows us to take a source separation approach which enables exemplar-based feature enhancement as well as missing data mask estimation. We evaluate the performance of these exemplar-based methods in connected digit recognition on the AURORA-2 database. Our results show that the hybrid system performed substantially better than source separation or missing data mask estimation at lower signal-to-noise ratios (SNRs), achieving up to 57.1% accuracy at SNR = -5 dB. Although not as effective as two baseline recognizers at higher SNRs, the novel approach offers a promising direction of future research on exemplar-based ASR.

In this paper we show that microphones that capture the bone-conducted voice can be used to improve Automatic Speech Recognition in noisy environments. These microphones exhibit good noise rejection properties and their signals are therefore good indications of speech activity, even in very noisy conditions. We conducted experiments where we used a throat microphone signal as a Voice Activity Detection (VAD) input signal and found that recognition accuracies in non-stationary noise improve significantly compared to when VAD is executed on a conventional microphone signal.

Partially supported by TUBITAK grant 106E071. We construct a robust localization framework to handle noisy measurements in wireless sensor networks. Traditionally many approaches employ the distance information gathered from ranging devices of the sensor nodes to achieve localization. However the measurements of these devices may contain noise both as hardware noise and as environmental noise due to the employment conditions of the network. It Is necessary to provide a general framework that handles such a noise in data and yet still be applicable within several localization algorithms. In order to handle noise in distance measurements, our framework utilizes convex constraints and confidence intervals of a random variable. At the end of the localization process nodes are assigned to a set of feasible regions with corresponding probabilities. The accuracy of the localization can be adjusted and the framework can easily be embedded to work within previously suggested localization algorithms. Türkiye Bilimsel ve Teknolojik Araştırma Kurumu (TÜBİTAK) Publisher's Version WOS:000275024200125

In this paper, a variance constrained filtering problem is considered for systems with both non-Gaussian noises and polytopic uncertainty. A novel filter is developed to estimate the systems states based on the current observation and known deterministic input signals. A free parameter is introduced in the filter to handle the uncertain input matrix in the known deterministic input term. In addition, unlike the existing variance constrained filters, which are constructed by the previous observation, the filter is formed from the current observation. A time-varying linear matrix inequality (LMI) approach is used to derive an upper bound of the state estimation error variance. The optimal bound is obtained by solving a convex optimisation problem via Semi-Definite Programming (SDP) approach. Simulation results are provided to demonstrate the effectiveness of the proposed method.

This thesis dissertation proposes, as a first step, a detailed introduction to the automatic speech recognition with missing data supported by many bibliographic references. It is shown that the estimation of masks is a crucial step. Indeed, the quality of the estimated masks determines the performance of the recognition system. Improving the reliability of masks is thus an important issue. In a second step, new investigations in the field of Bayesian missing data mask estimation are presented. I propose first new mask models to model dependencies between the masks of different coefficients of a signal. These models are evaluated and compared to a reference model. The results are presented in terms of error of masks, as well as recognition rate. The results show that these dependencies contribute to improving the recognition rate and stress the importance of the temporal context of a mask. Second, I introduce a new missing data mask definition: the masks of contribution. These new masks are evaluated compared to masks commonly used, based on the SNR thresholding. I show how the decoding algorithm can be improved with such a mask definition by refining the likelihood marginalization intervals. The assessment, in the context of data marginalization and in the presence of a stationary noise, shows that the intervals are considerably reduced resulting in a significant improvement of the recognition rate., Ce mémoire propose, dans un premier temps, une introduction détaillée de la reconnaissance automatique de la parole avec données manquantes appuyée par de nombreuses références bibliographiques. Il est montré que l'estimation de masques constitue une étape cruciale. En effet, la qualité des masques estimés conditionne les performances du système de reconnaissance. L'amélioration de la fiabilité des masques constitue donc un enjeu important. Dans un second temps, les travaux menés dans le cadre de l'estimation bayésienne des masques de données manquantes sont présentés. D'une part je propose de nouveaux modèles de masques permettant de modéliser les dépendances entre les masques de différents coefficients d'un signal. Ces modèles sont évalués comparativement à un modèle de référence. Les résultats sont présentés en termes d'erreur de masques ainsi qu'en taux de reconnaissance. Les résultats montrent que ces dépendances contribuent à améliorer les taux de reconnaissance et soulignent l'importance du contexte temporel d'un masque. Je présente, dans un second temps, une nouvelle définition de masque: les masques de contribution. Ces nouveaux masques sont évalués comparativement aux masques usuellement utilisés, fondés sur le seuillage du SNR. Je montre que cette nouvelle définition permet d'améliorer l'algorithme de décodage en affinant les intervalles de marginalisation. L'évaluation, dans le cadre de la marginalisation de données et en présence d'un bruit stationnaire, montrent que les intervalles sont considérablement réduits entraînant une nette amélioration des taux de reconnaissance.

We present in this paper a novel statistical based focalized reconstruction method for the underdetermined EEG (electroencephalogram) inverse problem. The algorithm is based on the representation of non-Gaussian distributions as an infinite mixture of Gaussians (IMG) and relies on an iterative procedure consisting out of alternated variance estimation/linear inversion operations. By taking into account noise statistics, it performs implicit spurious data rejection and produces robust focalized solutions allowing for straightforward discrimination of active/non-active brain regions. We apply the proposed reconstruction procedure to average evoked potentials EEG data and compare the reconstruction results with the corresponding known physiological responses.

Colloque avec actes et comité de lecture. internationale.; International audience; We present a novel noise compensation architecture which makes no assumptions on how the noise sources alter the speech data and which do not rely on clean speech models. Rather, this new architecture makes the (realistic) assumption that speech databases recorded under different background noise conditions are available. Its main principle is to process individually each database and to construct a parametric representation which describes the variation of acoustic models w.r.t. noise models. This representation is then used during recognition to estimate the acoustic models in the new environment. We evaluate the performance of this new compensation scheme on a connected digits recognition task and show that it can perform significantly better than multi-conditions training, which is the most widely used technique in these kind of scenarios.

Article dans revue scientifique avec comité de lecture. internationale.; International audience; The robustness of automatic speech recognition systems to noise is still a problem, especially for small footprint systems. This paper addresses the problem of noise robustness using model compensation methods. Such algorithms are already available, but their complexity is usually high. An often referenced method for achieving noise robustness is Parallel Model Combination (PMC). Several algorithms have been proposed to develop more computationally efficient methods than PMC. For example, Jacobian adaptation approximates PMC with a linear transformation function in the cepstral domain. However, the Jacobian approximation is valid only for test environments that are close to the training conditions whereas, in real test conditions, the mismatch between the test and training environments is usually large. In this paper, we propose two methods, respectively called static and dynamic alpha-Jacobian adaptation (or alpha-JAC), to compute new linear approximations of PMC for realistic test environments. We further extend both algorithms to compensate for additive and convolutional noise and we derive the corresponding non-linear algorithm that is approximated. All these algorithms are experimentally compared in important mismatch conditions. As compared to Jacobian adaptation, improvements are observed with both static and dynamic alpha-Jacobian adaptation.

This report proposes a mathematical and algorithmic study of the Lambertian SFS problem for orthographic and pinhole cameras. Our approach is based upon the notion of viscosity solutions of Hamilton-Jacobi equations. This approach provides a mathematical framework in which we can prove the well-posedness of the problem (proof of the existence of a solution and characterization of all solutions). This mathematical approach allows also to prove the correctness of our methods. In particular, we describe a simple monotonous stability condition for the studied decentered schemes and we prove the convergence of their solutions toward the viscosity solution of the associated Hamilton-Jacobi-Bellman equation. Also, we show that this theory naturally applies to the SFS problems. Our work extends previous work in the SFS area in three directions. First, it models the camera both as orthographic and as perspective (pinhole), i.e whereas most authors assume an orthographic projection (see for a panorama of the SFS problem up to 1989 and for more recent surveys); thus we extend the applicability of shape from shading methods to more realistic acquisition models. In particular it extends the work of and . Also, by introducing a «generic» Hamiltonian, we work in a general framework allowing to deal with both models, thereby simplifying the formalization of the problem. Second, it gives some novel mathematical formulations of this problem yielding new partial differential equations. Results about the existence and uniqueness of their solution are also obtained. Third, it allows us to come up with two new generic algorithms for computing numerical approximations of the "continuous" solution (of the «generic SFS problem») as well as a proof of their convergence toward that solution. Moreover, our two generic algorithms are able to deal with discontinuous images as well as images containing black shadows. Also, one of the algorithms we propose in this report, seems to be the most effective iterative algorithm of the SFS literature. From a more general viewpoint, our numerical results follow from a new method for solving Hamilton-Jacobi-Bellman equations. We propose two decentered finite difference schemes. We detail the proofs of the stability and the consistency of these schemes, and the proof of the convergence of their associated algorithms.

Colloque avec actes et comité de lecture. internationale.; International audience; We study in this work the Missing Data Recognition (MDR) framework applied to a large vocabulary continuous speech recognition (LVCSR) task with cepstral models when the speech signal is corrupted by musical noise. We do not propose a full system that solves this difficult problem, but we rather present some of the issues involved and study some possible solutions to them. We focus in this work on the issues concerning the application of masks to cepstral models. We further identify possible errors and study how some of them affect the performances of the system.

International audience; This paper analyzes the suitability of various multiple access schemes for a 60 GHz mobile ad hoc network. The investigated techniques include multi-carrier modulation (OFDM-TDMA), the spectrum spreading method (DS-CDMA) and the combined schemes (MC-CDMA, MC-DS-CDMA) benefiting from the advantages of both techniques. The comparison between these different multiple access schemes is focussed on two aspects: (i) sensitivity to phase noise; the large available bandwidth around 60 GHz is very suitable for transmission of high data rate in indoor environments, but one issue in the 60 GHz band is the design of oscillators with moderate phase noise; (ii) robustness to imperfect power control; ad hoc networks consist of a set of mobile terminals communicating among themselves without any central controller; radio resource management has to be conducted in a distributed way and the power control is inherently imperfect.

Colloque avec actes et comité de lecture. internationale.; International audience; Mel-frequency cepstral coefficients (MFCC) are the most widely used features in current speech recognition systems. However, they have a poor physical interpretation and they do not lie in the frequency domain. Frequency filtering (FF) is a technique that has been recently developed to design frequency-localized speech features that perform similar to MFCC in terms of recognition performances. Motivated by our desire to build time-frequency speech models, we wanted to use the FF technique as front-end. However, when evaluating FF on the Aurora-3 database we found some discrepancies in the highly mismatch case. This led us to put FF in another perspective: the wavelet transform. By doing so, we were able to explain the discrepancies and to achieve significant improvements in recognition in the highly mismatch case.

Colloque avec actes et comité de lecture. nationale.; National audience; We present a new predictive compensation scheme which makes no assumption on how the noise sources alter the speech data and which do not rely on clean speech models. Rather, this new scheme makes the (realistic) assumption that speech databases recorded under different background noise conditions are available. The philosophy of this scheme is to process these databases in order to build a "tool" which will allow it to handle new noise conditions in a robust way. We evaluate the performances of this new compensation scheme on a connected digits recognition task and show that it can perform significantly better than multi-conditions training, which is the most widely used techniques in these kind of scenarios.

International audience; In this paper we develop an algorithm to improve the accuracy of the estimation of the direction of arrival of the wide-band sources. It is well known that when the noise cross-spectral matrix is unknown, these estimates may be grossly inaccurate. Using both the fourth order cumulant for suppression of the Gaussian noise, the transformation matrices for estimating the coherent signal subspace and a noneigenvector algorithm a robust method for the source characterisation problem in the presence of noise with an unknown cross-spectral matrix is developed. We show that the performance of bearing estimation algorithms improves substantially when our robust algorithm is used. Simulation results are presented for the unknown noise spectral matrix.

In this paper, a new edge detector using a multilayer neural network, called a neural edge detector (NED), is proposed for detecting the desired edges clearly from noisy images. The NED is a supervised edge detector: through training the NED with a set of input images and desired edges, it acquires the function of a desired edge detector. The experiments on the NED to detect the edges from noisy test images and noisy natural images were performed. By comparative evaluation with the conventional edge detectors, the following has been demonstrated: the NED is robust against noise; the NED can detect clear continuous edges from the noisy images; and the performance of the NED is the highest in terms of similarity to the desired edges.

This thesis presents a study of alternative speech feature extraction methods aimed at increasing robustness of automatic speech recognition (ASR) against additive background noise. Spectral peak positions of speech signals remain practically unchanged in presence of additive background noise. Thus, it was expected that emphasizing spectral peak positions in speech feature extraction would result in improved noise robustness of ASR systems. If frequency subbands are properly chosen, dominant subband frequencies can serve as reasonable estimates of spectral peak positions. Thus, different methods for incorporating dominant subband frequencies into speech feature vectors were investigated in this study. To begin with, two earlier proposed feature extraction methods that utilize dominant subband frequency information were examined. The first one uses zero-crossing statistics of the subband signals to estimate dominant subband frequencies, while the second one uses subband spectral centroids. The methods were compared with the standard MFCC feature extraction method on two different recognition tasks in various background conditions. The first method was shown to improve ASR performance on both recognition tasks at sufficiently high noise levels. The improvement was, however, smaller on the more complex recognition task. The second method, on the other hand, led to some reduction in ASR performance in all testing conditions. Next, a new method for incorporating subband spectral centroids into speech feature vectors was proposed, and was shown to be considerably more robust than the standard MFCC method on both ASR tasks. The main difference between the proposed method and the zero-crossing based method is in the way they utilize dominant subband frequency information. It was shown that the performance improvement due to the use of dominant subband frequency information was considerably larger for the proposed method than for the ZCPA method, especially on the more complex recognition task. Finally, the computational complexity of the proposed method is two orders of magnitude lower than that of the zero-crossing based method, and of the same order of magnitude as the standard MFCC method. dr.ing. dr.ing.

International audience; In this paper a fusion algorithm for the detection of specular echoes blurred by coloured Gaussian noise is proposed. The phase and energy criteria, characteristic of these echoes, are handled separately. First, the detection capabilities of split spectrum (phase) and short time Fourier transform spectrogram (energy) are compared theoretically and experimentally. Second, the new algorithm based on mathematical morphology operators using both time-frequency representations is proposed. In the last part our method is shown to be more robust than absolute minimisation split spectrum processing.

International audience; In this study, we develop a robust method to improve the accuracy of the estimation of the direction-of-arrival (DOA) of wideband sources in the presence of noise with an unknown spectral matrix. It is well known that when the noise spectral matrix is unknown, those estimates may be grossly inaccurate. Our approach relies on noise subspace methods using the propagator operator. Finally, fast transformation matrices are presented which lead to the construction of a coherent signal subspace. The simulation results concerning the performance of the developed method are presented.

The performance of the existing speech recognition systems degrades rapidly in the presence of background noise. A novel representation of the speech signal, which is based on Linear Prediction of the One-Sided Autocorrelation sequence (OSALPC), has shown to be attractive for noisy speech recognition because of both its high recognition performance with respect to the conventional LPC in severe conditions of additive white noise and its computational simplicity. The aim of this work is twofold: (1) to show that OSALPC also achieves a good performance in a case of real noisy speech (in a car environment), and (2) to explore its combination with several robust similarity measuring techniques, showing that its performance improves by using cepstral liftering, dynamic features and multilabeling.

We consider three-dimensional cubic barcodes, consisting of smaller cubes, each built from one of two possible materials and carry one bit of information. To retrieve the information stored in the barcode, we measure a 2-D projection of the barcode using a penetrating wave such as X-rays, either using parallel-beam or cone-beam scanners from an unknown direction. We derive a theoretical representation of this scanning process and show that for a known barcode pose with respect to the scanner, the projection operator is linear and can be easily inverted. Moreover, we provide a method to estimate the unknown pose of the barcode from a single 2-D scan. We also propose coding schemes to correct errors and ambiguities in the reconstruction process. Finally, we test our designed barcode and reconstruction algorithms with several simulations, as well as a real-world barcode acquired with an X-ray cone-beam scanner, as a proof of concept.

Cepstral normalisation in automatic speech recognition is investigated in the context of robustness to additive noise. In this paper, it is argued that such normalisation leads naturally to a speech feature based on signal to noise ratio rather than absolute energy (or power). Explicit calculation of this SNR-cepstrum by means of a noise estimate is shown to have theoretical and practical advantages over the usual (energy based) cepstrum. The relationship between the SNR-cepstrum and the articulation index, known in psycho-acoustics, is discussed. Experiments are presented suggesting that the combination of the SNR-cepstrum with the well known perceptual linear prediction method can be beneficial in noisy environments.