Um software de reconhecimento de voz para português brasileiro

Descreve a implementação de um software de reconhecimento de voz para o Português Brasileiro. Dentre os objetivos do trabalho tem-se a construção de um sistema de voz contínua para grandes vocabulários, apto a ser usado em aplicações em tempo-real. São apresentados os principais conceitos e características de tais sistemas, além de todos os passos necessários para construção. Como parte desse trabalho foram produzidos e disponibilizados vários recursos: modelos acústicos e de linguagem, novos corpora de voz e texto. O corpus de texto vem sendo construído através da extração e formatação automática de textos de jornais na Internet. Além disso, foram produzidos dois corpora de voz, um baseado em audiobooks e outro produzido especificamente para simular testes em tempo-real. O trabalho também propõe a utilização de técnicas de adaptação de locutor para resolução de problemas de descasamento acústico entre corpora de voz. Por último, é apresentada uma interface de programação de aplicativos que busca facilitar a utilização do decodificador Julius. Testes de desempenho são apresentados, comparando os sistemas desenvolvidos e um software comercial.

Automatic aspect discrimination in data clustering

The attributes describing a data set may often be arranged in meaningful subsets, each of which corresponds to a different aspect of the data. An unsupervised algorithm (SCAD) that simultaneously performs fuzzy clustering and aspects weighting was proposed in the literature. However, SCAD may fail and halt given certain conditions. To fix this problem, its steps are modified and then reordered to reduce the number of parameters required to be set by the user. In this paper we prove that each step of the resulting algorithm, named ASCAD, globally minimizes its cost-function with respect to the argument being optimized. The asymptotic analysis of ASCAD leads to a time complexity which is the same as that of fuzzy c-means. A hard version of the algorithm and a novel validity criterion that considers aspect weights in order to estimate the number of clusters are also described. The proposed method is assessed over several artificial and real data sets.

Automatic clothes segmentation for soft biometrics

[EN]During the last decade, researchers have verified that clothing can provide information for gender recognition. However, before extracting features, it is necessary to segment the clothing region. We introduce a new clothes segmentation method based on the application of the GrabCut technique over a trixel mesh, obtaining very promising results for a close to real time system. Finally, the clothing features are combined with facial and head context information to outperform previous results in gender recognition with a public database.

An analysis of automatic gender classification

[EN]Different researches suggest that inner facial features are not the only discriminative features for tasks such as person identification or gender classification. Indeed, they have shown an influence of features which are part of the local face context, such as hair, on these tasks. However, object-centered approaches which ignore local context dominate the research in computational vision based facial analysis. In this paper, we performed an analysis to study which areas and which resolutions are diagnostic for the gender classification problem. We first demonstrate the importance of contextual features in human observers for gender classification using a psychophysical ”bubbles” technique.

Face recognition from a tabula rasa perspective

[EN]In this paper a system for face recognition from a tabula rasa (i.e. blank slate) perspective is described. A priori, the system has the only ability to detect automatically faces and represent them in a space of reduced dimension. Later, the system is exposed to over 400 different identities, observing its recognition performance evolution. The preliminary results achieved indicate on the one side that the system is able to reject most of unknown individuals after an initialization stage.

Oblique ionograms automatic scaling and eikonal based ray tracing

A method for automatic scaling of oblique ionograms has been introduced. This method also provides a rejection procedure for ionograms that are considered to lack sufficient information, depicting a very good success rate. Observing the Kp index of each autoscaled ionogram, can be noticed that the behavior of the autoscaling program does not depend on geomagnetic conditions. The comparison between the values of the MUF provided by the presented software and those obtained by an experienced operator indicate that the procedure developed for detecting the nose of oblique ionogram traces is sufficiently efficient and becomes much more efficient as the quality of the ionograms improves. These results demonstrate the program allows the real-time evaluation of MUF values associated with a particular radio link through an oblique radio sounding. The automatic recognition of a part of the trace allows determine for certain frequencies, the time taken by the radio wave to travel the path between the transmitter and receiver. The reconstruction of the ionogram traces, suggests the possibility of estimating the electron density between the transmitter and the receiver, from an oblique ionogram. The showed results have been obtained with a ray-tracing procedure based on the integration of the eikonal equation and using an analytical ionospheric model with free parameters. This indicates the possibility of applying an adaptive model and a ray-tracing algorithm to estimate the electron density in the ionosphere between the transmitter and the receiver An additional study has been conducted on a high quality ionospheric soundings data set and another algorithm has been designed for the conversion of an oblique ionogram into a vertical one, using Martyn's theorem. This allows a further analysis of oblique soundings, throw the use of the INGV Autoscala program for the automatic scaling of vertical ionograms.

Deep learning for computer vision: a comparison between convolutional neural networks and hierarchical temporal memories on object recognition tasks

In recent years, Deep Learning techniques have shown to perform well on a large variety of problems both in Computer Vision and Natural Language Processing, reaching and often surpassing the state of the art on many tasks. The rise of deep learning is also revolutionizing the entire field of Machine Learning and Pattern Recognition pushing forward the concepts of automatic feature extraction and unsupervised learning in general. However, despite the strong success both in science and business, deep learning has its own limitations. It is often questioned if such techniques are only some kind of brute-force statistical approaches and if they can only work in the context of High Performance Computing with tons of data. Another important question is whether they are really biologically inspired, as claimed in certain cases, and if they can scale well in terms of "intelligence". The dissertation is focused on trying to answer these key questions in the context of Computer Vision and, in particular, Object Recognition, a task that has been heavily revolutionized by recent advances in the field. Practically speaking, these answers are based on an exhaustive comparison between two, very different, deep learning techniques on the aforementioned task: Convolutional Neural Network (CNN) and Hierarchical Temporal memory (HTM). They stand for two different approaches and points of view within the big hat of deep learning and are the best choices to understand and point out strengths and weaknesses of each of them. CNN is considered one of the most classic and powerful supervised methods used today in machine learning and pattern recognition, especially in object recognition. CNNs are well received and accepted by the scientific community and are already deployed in large corporation like Google and Facebook for solving face recognition and image auto-tagging problems. HTM, on the other hand, is known as a new emerging paradigm and a new meanly-unsupervised method, that is more biologically inspired. It tries to gain more insights from the computational neuroscience community in order to incorporate concepts like time, context and attention during the learning process which are typical of the human brain. In the end, the thesis is supposed to prove that in certain cases, with a lower quantity of data, HTM can outperform CNN.

The Role of Features and Context in Recognition of Novel Melodies

WE INVESTIGATED HOW WELL STRUCTURAL FEATURES such as note density or the relative number of changes in the melodic contour could predict success in implicit and explicit memory for unfamiliar melodies. We also analyzed which features are more likely to elicit increasingly confident judgments of "old" in a recognition memory task. An automated analysis program computed structural aspects of melodies, both independent of any context, and also with reference to the other melodies in the testset and the parent corpus of pop music. A few features predicted success in both memory tasks, which points to a shared memory component. However, motivic complexity compared to a large corpus of pop music had different effects on explicit and implicit memory. We also found that just a few features are associated with different rates of "old" judgments, whether the items were old or new. Rarer motives relative to the testset predicted hits and rarer motives relative to the corpus predicted false alarms. This data-driven analysis provides further support for both shared and separable mechanisms in implicit and explicit memory retrieval, as well as the role of distinctiveness in true and false judgments of familiarity.

Automatic Food Intake Assessment Using Camera Phones

Obesity is becoming an epidemic phenomenon in most developed countries. The fundamental cause of obesity and overweight is an energy imbalance between calories consumed and calories expended. It is essential to monitor everyday food intake for obesity prevention and management. Existing dietary assessment methods usually require manually recording and recall of food types and portions. Accuracy of the results largely relies on many uncertain factors such as user's memory, food knowledge, and portion estimations. As a result, the accuracy is often compromised. Accurate and convenient dietary assessment methods are still blank and needed in both population and research societies. In this thesis, an automatic food intake assessment method using cameras, inertial measurement units (IMUs) on smart phones was developed to help people foster a healthy life style. With this method, users use their smart phones before and after a meal to capture images or videos around the meal. The smart phone will recognize food items and calculate the volume of the food consumed and provide the results to users. The technical objective is to explore the feasibility of image based food recognition and image based volume estimation. This thesis comprises five publications that address four specific goals of this work: (1) to develop a prototype system with existing methods to review the literature methods, find their drawbacks and explore the feasibility to develop novel methods; (2) based on the prototype system, to investigate new food classification methods to improve the recognition accuracy to a field application level; (3) to design indexing methods for large-scale image database to facilitate the development of new food image recognition and retrieval algorithms; (4) to develop novel convenient and accurate food volume estimation methods using only smart phones with cameras and IMUs. A prototype system was implemented to review existing methods. Image feature detector and descriptor were developed and a nearest neighbor classifier were implemented to classify food items. A reedit card marker method was introduced for metric scale 3D reconstruction and volume calculation. To increase recognition accuracy, novel multi-view food recognition algorithms were developed to recognize regular shape food items. To further increase the accuracy and make the algorithm applicable to arbitrary food items, new food features, new classifiers were designed. The efficiency of the algorithm was increased by means of developing novel image indexing method in large-scale image database. Finally, the volume calculation was enhanced through reducing the marker and introducing IMUs. Sensor fusion technique to combine measurements from cameras and IMUs were explored to infer the metric scale of the 3D model as well as reduce noises from these sensors.

Segmentation and Recognition of Multi-Food Meal Images for Carbohydrate Counting

In this paper, we propose novel methodologies for the automatic segmentation and recognition of multi-food images. The proposed methods implement the first modules of a carbohydrate counting and insulin advisory system for type 1 diabetic patients. Initially the plate is segmented using pyramidal mean-shift filtering and a region growing algorithm. Then each of the resulted segments is described by both color and texture features and classified by a support vector machine into one of six different major food classes. Finally, a modified version of the Huang and Dom evaluation index was proposed, addressing the particular needs of the food segmentation problem. The experimental results prove the effectiveness of the proposed method achieving a segmentation accuracy of 88.5% and recognition rate equal to 87%

Implicit memory for the content but not the speaker of sleep-played messages

We presented 28 sentences uttered by 28 unfamiliar speakers to sleeping participants to investigate whether humans can encode new verbal messages, learn voices of unfamiliar speakers, and form associations between speakers and messages during EEG-defined deep sleep. After waking, participants performed three tests which assessed the unconscious recognition of sleep-played speakers, messages, and speaker-message associations. Recognition performance in all tests was at chance level. However, response latencies revealed implicit memory for sleep-played messages but neither for speakers nor for speaker-message combinations. Only participants with excellent implicit memory for sleep-played messages also displayed implicit memory for speakers but not speaker-message associations. Hence, deep sleep allows for the semantic encoding of novel verbal messages.

Automatic Understanding of ATC Speech: Study of Prospectives and Field Experiments for Several Controller Positions

Although there has been a lot of interest in recognizing and understanding air traffic control (ATC) speech, none of the published works have obtained detailed field data results. We have developed a system able to identify the language spoken and recognize and understand sentences in both Spanish and English. We also present field results for several in-tower controller positions. To the best of our knowledge, this is the first time that field ATC speech (not simulated) is captured, processed, and analyzed. The use of stochastic grammars allows variations in the standard phraseology that appear in field data. The robust understanding algorithm developed has 95% concept accuracy from ATC text input. It also allows changes in the presentation order of the concepts and the correction of errors created by the speech recognition engine improving it by 17% and 25%, respectively, absolute in the percentage of fully correctly understood sentences for English and Spanish in relation to the percentages of fully correctly recognized sentences. The analysis of errors due to the spontaneity of the speech and its comparison to read speech is also carried out. A 96% word accuracy for read speech is reduced to 86% word accuracy for field ATC data for Spanish for the "clearances" task confirming that field data is needed to estimate the performance of a system. A literature review and a critical discussion on the possibilities of speech recognition and understanding technology applied to ATC speech are also given.

Statistical Machine Learning for Automatic Assessment of Physical Activity Intensity Using Multi-axial Accelerometry and Heart Rate

This work explores the automatic recognition of physical activity intensity patterns from multi-axial accelerometry and heart rate signals. Data collection was carried out in free-living conditions and in three controlled gymnasium circuits, for a total amount of 179.80 h of data divided into: sedentary situations (65.5%), light-to-moderate activity (17.6%) and vigorous exercise (16.9%). The proposed machine learning algorithms comprise the following steps: time-domain feature definition, standardization and PCA projection, unsupervised clustering (by k-means and GMM) and a HMM to account for long-term temporal trends. Performance was evaluated by 30 runs of a 10-fold cross-validation. Both k-means and GMM-based approaches yielded high overall accuracy (86.97% and 85.03%, respectively) and, given the imbalance of the dataset, meritorious F-measures (up to 77.88%) for non-sedentary cases. Classification errors tended to be concentrated around transients, what constrains their practical impact. Hence, we consider our proposal to be suitable for 24 h-based monitoring of physical activity in ambulatory scenarios and a first step towards intensity-specific energy expenditure estimators