Aldebaro Klautau

On nearest-neighbor error-correcting output codes with application to all-pairs multiclass support vector machines

A common way of constructing a multiclass classifier is by combining the outputs of several binary ones, according to an error-correcting output code (ECOC) scheme. The combination is typically done via a simple nearest-neighbor rule that finds the class that is closest in some sense to the outputs of the binary classifiers. For these nearest-neighbor ECOCs, we improve existing bounds on the error rate of the multiclass classifier given the average binary distance. The new bounds provide insight into the one-versus-rest and all-pairs matrices, which are compared through experiments with standard datasets. The results also show why elimination (also known as DAGSVM) and Hamming decoding often achieve the same accuracy.

Semiblind Spectrum Balancing for DSL

Digital subscriber lines (DSLs) technology is vastly used for high-speed data transmission. Crosstalk is one of the main problems in DSL. The research community has done extensive work on how to optimize spectrum allocation across DSL frequencies and mitigate crosstalk, a subject that has been called dynamic spectrum management (DSM). This text presents a novel DSM solution, referred to as semiblind spectrum balancing (2SB). 2SB performs an optimization process with a virtual line, a fictitious line that represents to each user the damage it causes for all other users. With the aid of message exchanges between modems and a central agent, the method will adjust the virtual lines parameters so that it represents the real crosstalk scenario in the binder. In this paper, we provide the conditions for how such a situation can be achieved and show that it can do so with semicentralization, low complexity and limited crosstalk channel information-only the ratios between crosstalk channels are necessary. Performance is very close to optimal.

Line Topology Identification Using Multiobjective Evolutionary Computation

The broadband capacity of the twisted-pair lines strongly varies within the copper access network. It is therefore important to assess the ability of a digital subscriber line (DSL) to support the DSL services prior to deployment. This task is handled by the line qualification procedures, where the identification of the line topology is an important part. This paper presents a new method, denoted topology identification via model-based evolutionary computation (TIMEC), for line topology identification, where either one-port measurements or both one- and two-port measurements are utilized. The measurements are input to a model-based multiobjective criterion that is minimized by a genetic algorithm to provide an estimate of the line topology. The inherent flexibility of TIMEC enables the incorporation of a priori information, e.g., the total line length. The performance of TIMEC is evaluated by computer simulations with varying degrees of information. Comparison with a state-of-art method indicates that TIMEC achieves better results for all the tested lines when only one-port measurements are used. The results are improved when employing both one- and two-port measurements. If a rough estimate of the total length is also used, near-perfect estimation is obtained for all the tested lines.

Spectrum Balancing Algorithms for Power Minimization in DSL Networks

Spectrum balancing (SB) techniques optimize transmission and can significantly improve digital subscriber lines (DSL) services. In the literature, the DSL system optimization is typically formulated as a rate maximization problem. However, there is an increasing interest in minimizing the considerable amount of power consumed by telecommunication networks. Few works in the SB literature have explored algorithms for power minimization. It is known that some existing solutions for rate maximization can be converted into power minimization algorithms. This relation has not been fully explored and, consequently, the area lacks results regarding what can be achieved with power minimization SB algorithms. This work aims to diminish this gap. First, the equivalence between rate maximization and power minimization problems is formalized. Second, extended versions of some rate maximization SB algorithms are proposed for power minimization purposes and evaluated through simulations. In addition, the power-usage capabilities and convergence characteristics of each extended SB algorithms is discussed.

Free tools and resources for Brazilian Portuguese speech recognition

An automatic speech recognition system has modules that depend on the language and, while there are many public resources for some languages (e.g., English and Japanese), the resources for Brazilian Portuguese (BP) are still limited. This work describes the development of resources and free tools for BP speech recognition, consisting of text and audio corpora, phonetic dictionary, grapheme-to-phone converter, language and acoustic models. All of them are publicly available and, together with a proposed application programming interface, have been used for the development of several new applications, including a speech module for the OpenOffice suite. Performance tests are presented, comparing the developed BP system with a commercial software. The paper also describes an application that uses synthesis and speech recognition together with a natural language processing module dedicated to statistical machine translation. This application allows the translation of spoken conversations from BP to English and vice versa. The resources make easier the adoption of BP speech technologies by other academic groups and industry.

New approach for T-wave end detection on electrocardiogram: Performance in noisy conditions

BackgroundThe detection of T-wave end points on electrocardiogram (ECG) is a basic procedure for ECG processing and analysis. Several methods have been proposed and tested, featuring high accuracy and percentages of correct detection. Nevertheless, their performance in noisy conditions remains an open problem.MethodsA new approach and algorithm for T-wave end location based on the computation of Trapeziums areas is proposed and validated (in terms of accuracy and repeatability), using signals from the Physionet QT Database. The performance of the proposed algorithm in noisy conditions has been tested and compared with one of the most used approaches for estimating the T-wave end point: the method based on the threshold on the first derivative.ResultsThe results indicated that the proposed approach based on Trapeziums areas outperformed the baseline method with respect to accuracy and repeatability. Also, the proposed method is more robust to wideband noise.ConclusionsThe trapezium-based approach has a good performance in noisy conditions and does not rely on any empirical threshold. It is very adequate for use in scenarios where the levels of broadband noise are significant.

A Front End for Discriminative Learning in Automatic Modulation Classification

This work presents a novel method for automatic modulation classification based on discriminative learning. The features are the ordered magnitude and phase of the received symbols at the output of the matched filter. The results using the proposed front end and support vector machines are compared to other techniques. Frequency offset is also considered and the results show that in this condition the new method significantly outperforms two cumulant-based classifiers.

An open-source speech recognizer for Brazilian Portuguese with a windows programming interface

This work is part of the effort to develop a speech recognition system for Brazilian Portuguese. The resources for the training and test stages of this system, such as corpora, pronunciation dictionary, language and acoustic models, are publicly available. Here, an application programming interface is proposed in order to facilitate using the open-source Julius speech decoder. Performance tests are presented, comparing the developed systems with a commercial software.

Optimizing Power Normalization for G.fast Linear Precoder by Linear Programming

The use of vectoring for crosstalk cancellation in the new ITU-T G.fast standard for next generation DSL systems becomes essential for efficient utilization of the extended bandwidth (up to 200 MHz). In VDSL2 (up to 30 MHz), a zero-forcing-based linear precoder is used in downstream which approaches single-line performance. However, at high frequencies, the linear precoder may amplify the signal power substantially since the crosstalk channel is much stronger than at lower frequencies. Performance could be significantly degraded by power normalization to keep the PSD below the mask. In this work, we extended a per-line power normalization scheme by linear programming (LP) optimization. By simulations using measured cable data it is shown how the LP-based scheme further improves the linear precoder and it is also capable of balancing the data rate between lines. Further, the simulations also show the non-linear Tomlinson-Harashima precoder performs better than the linear precoders.

Simple and Causal Copper Cable Model Suitable for G.fast Frequencies

G.fast is a new standard from the International Telecommunication Union, which targets 1 Gb/s over short copper loops using frequencies up to 212 MHz. This new technology requires accurate parametric cable models for simulation, design, and performance evaluation tests. Some existing copper cable models were designed for the very high speed digital subscriber line spectra, i.e., frequencies up to 30 MHz, and adopt assumptions that are violated when the frequency range is extended to G.fast frequencies. This paper introduces a simple and causal cable model that is able to accurately characterize copper loops composed by single or multiple segments, in both frequency and time domains. Results using G.fast topologies show that, apart from being accurate, the new model is attractive due to its low computational cost and closed-form expressions for fitting its parameters to measurement data.