Abstract
It is well known that the performance of Machine Learning techniques, notably when applied to Computer Vision (CV), depends heavily on the amount and quality of the training data set. However, large data sets lead to time-consuming training loops and, in many situations, are difficult or even impossible to create. Therefore, there is a need for solutions to reduce their size while ensuring good levels of performance, i.e., solutions that obtain the best tradeoff between the amount/quality of training data and the model's performance. This paper proposes a dataset reduction approach for training data used in Deep Learning methods in Facial Recognition (FR) problems. We focus on maximizing the variability of representations for each subject (person) in the training data, thus favoring quality instead of size. The main research questions are: 1) Which facial features better discriminate different identities? 2) Will it be possible to significantly reduce the training time without compromising performance? 3) Should we favor quality over quantity for very large datasets in FR? This analysis uses a pipeline to discriminate a set of features suitable for capturing the diversity and a cluster-based sampling to select the best images for each training subject, i.e., person. Results were obtained using VGGFace2 and Labeled Faces in the Wild (for benchmarking) and show that, with the proposed approach, a data reduction is possible while ensuring similar levels of accuracy.

Abstract
This paper proposes a method to avoid using a CORDIC or external memory to process the steering vectors to calculate the pseudospectrum of correlation-based beamforming algorithms. We show that if we decompose the steering vector equation, the size of the matrix to be saved in memory becomes independent of the antenna array size. Besides, the amount of data needed is small enough to be saved in the internal block RAMs of the FPGA SoC. Besides, this method greatly reduces the number of memory accesses, by offloading some processing to hardware, while keeping the frequency at 300MHz with a precision of 0.25 degrees. Finally, we show that this approach is scalable since the complexity grows logarithmically for bigger arrays, and the symmetry in the matrices obtained allows even more compact data.

Abstract
Many natural signals exhibit a quasi-periodic behavior and are conveniently modeled as a combination of several harmonic sinusoids whose relative frequencies, magnitudes and phases vary with time. The waveform shape of those signals reflects important physical phenomena underlying their generation, which requires that those parameters be accurately estimated and modeled. In the literature, accurate phase estimation and modeling has received much less research effort than frequency estimation, or magnitude estimation. First, this paper addresses accurate DFT-based phase estimation of individual sinusoids in six scenarios involving two DFT-based filter banks and three different windows. It is shown that bias in phase estimation is less than 1E-3 radians when the SNR is equal to or larger than 2.5 dB. Taking as a reference the Cramér-Rao Lower Bound, it is shown that one particular window offers a performance of practical interest by approximating better the CRLB when signal conditions are favorable, and by minimizing the performance deviation when signal conditions are adverse. Second, this paper explains how a shift-invariant phase-related feature can be devised that characterizes harmonic phase structure, which motivates a signal processing paradigm that greatly simplifies parametric modeling, transformation and synthesis of harmonics signals, in addition to facilitating the understanding and reverse engineering of the phasegram. Theory and results are discussed in a reproducible perspective using dedicated experiments that are supported with code allowing not only to replicate figures and results in this paper, but also to expand research.

Abstract
Many natural signals exhibit quasi-periodic behaviors and are conveniently modeled as combinations of several harmonic sinusoids whose relative frequencies, magnitudes, and phases vary with time. The waveform shapes of those signals reflect important physical phenomena underlying their generation, requiring those parameters to be accurately estimated and modeled. In the literature, accurate phase estimation and modeling have received significantly less attention than frequency or magnitude estimation. This paper first addresses accurate DFT-based phase estimation of individual sinusoids across six scenarios involving two DFT-based filter banks and three different windows. It has been shown that bias in phase estimation is less than 0.001 radians when the SNR is equal to or larger than 2.5 dB. Using the Cram & eacute;r-Rao lower bound as a reference, it has been demonstrated that one particular window offers performance of practical interest by better approximating the CRLB under favorable signal conditions and minimizing performance deviation under adverse conditions. This paper describes the development of a shift-invariant phase-related feature that characterizes the harmonic phase structure. This feature motivates a new signal processing paradigm that greatly simplifies the parametric modeling, transformation, and synthesis of harmonic signals. It also aids in understanding and reverse engineering the phasegram. The theory and results are discussed from a reproducible perspective, with dedicated experiments supported by code, allowing for the replication of figures and results presented in this paper and facilitating further research.

Abstract
The phase response of all-pole (AP) models is known to be non-linear and highly dependent on the frequency response magnitude. The objective and perceptual impact of the group delay of AP models in the synthesis of vowel sounds has not been thoroughly addressed in the literature. In this paper, we use a dedicated frequency-domain framework so as to i) synthesize a plausible glottal excitation setting the ground-truth for the harmonic phase structure and replicating the fundamental frequency contour of natural vowels, ii) synthesize realistic vowel sounds through all-zero (AZ) and all-pole (AP) models sharing the same frequency response magnitude, and iii) assess the objective and perceptual impact of the group delay of AP models taking as a reference natural vowels and, in particular, the ground-truth harmonic phase structure of the glottal excitation. Our findings emphasize that the non-linear phase characteristics of AP models degrade the harmonic phase structure of synthetic vowels significantly beyond what is found in natural vowels, however, that is not always clearly audible.

Abstract