Papers by Jugurta Montalvão
TEMA (São Carlos), 2014
In this paper, the seminal method proposed by Abraham Lempel and Jacob Ziv, aimed at the complexi... more In this paper, the seminal method proposed by Abraham Lempel and Jacob Ziv, aimed at the complexity analysis of sequences of symbols, was modified to compare similarities between two sequences. This modification allowed the creation of a new criterion which can replace likelihood in some pattern recognition applications. Moreover, to allow for analysis and comparison of multivariate continuously valued patterns, we also present a simple adaptation of the Lempel-Ziv's method to time-sampled signals. To illustrate the usefulness of these proposed tools, two sets of experimental results are presented, namely: one on speaker identity verification (biometrics) and another on healthcare signal detection. Both experiments yield promising performances. Moreover, as compared to a conventional pattern recognition method, the new approach provided better performances in terms of Equal Error Ratio in speaker verification experiments.
A new blind equalization algorithm for digital communication systems is presented. This algorithm... more A new blind equalization algorithm for digital communication systems is presented. This algorithm is based on the adjusting of a linear equalizer in such a way that the probability density function (PDF) of its output matches a parametric target function. A link between the proposed cost function and that used by the Constant Modulus Algorithm is also pointed out. Some simulation results are presented and compared to that provided by the Godard's equalizer.
The use of a nonlinear structure of filtering for blind equalization is presented. The structure ... more The use of a nonlinear structure of filtering for blind equalization is presented. The structure neural network-based is used in order to provide nonlinearity on the filter structure and the learning strategy is then divided in two stages. The Kullback-Leibler divergence is used as the base for the cost function of a self-organized rule and constant modulus criterion for the supervised one. Simulation results illustrate the performance of the strategy compared with classical ones for adaptive equalization. The results show that the proposed strategy outperforms even trained DFE for some cases of channels.
A family of nonlinear equalizers: Sub-optimal Bayesian classifiers
ABSTRACT
Une structure de filtrage non linéaire pour l'égalisation aveugle est présentée. Cette structure ... more Une structure de filtrage non linéaire pour l'égalisation aveugle est présentée. Cette structure est basée sur un réseau de neurones, ce qui permet l'inclusion de non linéarités dans la structure du filtre. D'autre part, la stratégie d'apprentissage du réseau est séparée en deux parties : une supervisée et l'autre auto-organisée. La divergence de Kullback-Leibler est utilisée comme base pour une fonction de coût d'une règle d'apprentissage auto-organisée, tandis que le critère du (( constant modulus )) est utilisé dans la partie supervisée. Les résultats des simulations comparent la performance de cette stratégie par rapport aux stratégies classiques d'égalisation adaptative. Les résultats montrent que, pour certains canaux, la stratégie proposée est plus performante que l'égaliseurà retour des décisions (DFE) supervisé.
An automatic method for validation and pruning of probability density function estimates is prese... more An automatic method for validation and pruning of probability density function estimates is presented. Densities are approximated by Normal Mixture Models (or Gaussian Mixture Models) whose Gaussian kernels are placed according to a modified Parzen method, whereas Gaussian dispersions are optimized through likelihood validation. Then, by assuming that resulting mixtures perfectly fit the actual densities (null hypothesis), a union of acceptance regions is defined and used for testing the null hypothesis. This hypothesis test yields an automatic kernel pruning procedure. Experimental results with both synthetic and real-world (biometric) data sets corroborate the expected improvement, in terms of regularization, of pruned density models.
Análise De Espectro Através Da Deteção De Eventos Acústicos Elementares No Plano Tempo-Frequência
An event-based method is presented as an alternative for spectral analysis of acoustic signals. T... more An event-based method is presented as an alternative for spectral analysis of acoustic signals. The targeted event is upward level-crossings, from which three essential information are taken: instantaneous frequency between consecutive events, corresponding maximum amplitude and instant of occurrence of each event. This approach is similar to EIH and ZCPA, in this conception, but innovative in terms of spectral estimation, since a Monte Carlo is used instead of histograms. Some illustrations of properties are discussed, such as the intrinsic silence suppression and nonlinear scaling of time, along with experimental illustration through isolated word detection. Resumo— Um método para análise espectral de sinais acústicos baseada em evento e apresentado. O evento alvo, nesse método e o cruzamento ascendente de nível, do qual são extraídas três informações essenciais: a frequência instantânea entre eventos consecutivos, a amplitude correspondentes, e o instante de ocorrência de cada ev...
Some signals are intrinsically symbolic, such as nucleotides, in a Deoxyri-bonucleic acid (DNA) m... more Some signals are intrinsically symbolic, such as nucleotides, in a Deoxyri-bonucleic acid (DNA) molecule, which is quite naturally seen as a mapping from (relative) position, x, to a letter that represents the nucleotides struc-ture — i.e. the finite set U = {A, G, T, C}. Representations for such signals are not naturally related to metric spaces. In the statistical literature, symbolic is usually termed categorical. Indeed, though symbolic signals may be named differently depending on professional jargon, it appears in Linguistics, Biology, Discrete event systems, Business Process Mining and Mathematics, etc. In this text, some selected tools related to symbolic processing are explained as plainly as possible manner.
A method for both speaker (Biometrics) and speech recognition is proposed. It was adjusted to obt... more A method for both speaker (Biometrics) and speech recognition is proposed. It was adjusted to obtain high performance under minimum training requirement: only one sample of each target-word, uttered by the target-subject. This minimum training requirement fits with remote health-care convenience, where long and tedious enrolment sessions, with patients or elderly people, are not welcome. In extract a maximum number of relevant cues from single samples, we improve two fragile steps in usual approaches based on Dynamic Time-Warping (DTW). First, we use a resampling strategy based on energy profile, associated to a greedy local search. Moreover, we replace usual scoring based on distances with an evidence accumulation score, in bits of information. Experimental results with two databases (more than 1000 samples) show that both strategies indeed provide improved performances. Resumo— Um método para reconhecimento de orador (biometria) e de comandos pronunciado e proposto neste artigo. S...
Journal of Communication and Information Systems, 1999
In this paper, we investigate some advantages and limitations of nonlinear architectures on adapt... more In this paper, we investigate some advantages and limitations of nonlinear architectures on adaptive equalization of dispersive and linear channels. The channel equalization problem is seen as one of classification, and we present some theoretical results which relate the zeros of the channel impulsive response and the geometrical dispersion of channel states. These theoretical results and some related consequences are illustrated by means of simple examples.
A pragmatic entropy and differential entropy estimator for small datasets
Journal of Communication and Information Systems, 2014
2011 International Joint Conference on Biometrics (IJCB), 2011
Breve descrição do desenvolvimento de um protótipo (hardware mais software) para o acionamento el... more Breve descrição do desenvolvimento de um protótipo (hardware mais software) para o acionamento eletrônico de dispositivo, onde o timbre de voz é usado como chave (ou assinatura) única dos usuários cadastrados. São sucintamente descritos os métodos retidos na extração de parâmetros e classificação dos sinais, ressaltando sempre o caráter fortemente didático do projeto, além do seu aspecto inovador no contexto estadual. Finalmente, uma medida de desempenho do sistema é sintetizada na forma de probabilidades estimadas de acertos e erros de classificação.
Abstract O resultado de experimentos com voluntários sobre a percepçao do ritmo da fala e do co... more Abstract O resultado de experimentos com voluntários sobre a percepçao do ritmo da fala e do contorno da entonaçao em uma sentença curta (3s de duraçao) é apresentado e confrontado a resultados obtidos através de sistemas automáticos de re-conhecimento de orador, ...
Probability density function estimation from limited data sets is a classical problem in pattern ... more Probability density function estimation from limited data sets is a classical problem in pattern recognition. In this paper we propose a reformulation of the well-known nonparametric Parzen method as a parametrically regularized Gaussian Mixture Model, from which we can easily estimate density contour level. As an application illustration to the proposed contour level estimator, we also address the Blind Source Separation problem through the analysis of contour level distortions in joint probability density functions. Finally, we use the proposed estimator to undo a nonlinear mixture of two images.
Computational vocalization recognition—an alternative to transect sampling in Tropical Forests
Percepção e Representação Vetorial De Timbre
Signal Features for Event Detection in Remote Health Care through Audio Signals
A very pragmatic approach for measuring memory (or inertia) in dynamic sequences of symbols (symb... more A very pragmatic approach for measuring memory (or inertia) in dynamic sequences of symbols (symbol dynamics) is proposed, where only intervals between coincidences of symbols along the sequence are taken into account in the process of estimating the Auto Mutual Information. The proposed method is studied using sequences of symbols obtained from two Markov sources, with two and nine states respectively. Results are compared to expected theoretical values of mutual information, as well as to histogram-based estimations with Miller's entropy bias compensation.
Uploads
Papers by Jugurta Montalvão