Técnicas de análisis, caracterización y detección de señales de voz en entornos acústicos adversos

Este trabajo de Tesis ha abordado el objetivo de dar robustez y mejorar la Detección de Actividad de Voz en entornos acústicos adversos con el fin de favorecer el comportamiento de muchas aplicaciones vocales, por ejemplo aplicaciones de telefonía basadas en reconocimiento automático de voz, aplicaciones en sistemas de transcripción automática, aplicaciones en sistemas multicanal, etc. En especial, aunque se han tenido en cuenta todos los tipos de ruido, se muestra especial interés en el estudio de las voces de fondo, principal fuente de error de la mayoría de los Detectores de Actividad en la actualidad. Las tareas llevadas a cabo poseen como punto de partida un Detector de Actividad basado en Modelos Ocultos de Markov, cuyo vector de características contiene dos componentes: la energía normalizada y la variación de la energía. Las aportaciones fundamentales de esta Tesis son las siguientes: 1) ampliación del vector de características de partida dotándole así de información espectral, 2) ajuste de los Modelos Ocultos de Markov al entorno y estudio de diferentes topologías y, finalmente, 3) estudio e inclusión de nuevas características, distintas de las del punto 1, para filtrar los pulsos de pronunciaciones que proceden de las voces de fondo. Los resultados de detección, teniendo en cuenta los tres puntos anteriores, muestran con creces los avances realizados y son significativamente mejores que los resultados obtenidos, bajo las mismas condiciones, con otros detectores de actividad de referencia. This work has been focused on improving the robustness at Voice Activity Detection in adverse acoustic environments in order to enhance the behavior of many vocal applications, for example telephony applications based on automatic speech recognition, automatic transcription applications, multichannel systems applications, and so on. In particular, though all types of noise have taken into account, this research has special interest in the study of pronunciations coming from far-field speakers, the main error source of most activity detectors today. The tasks carried out have, as starting point, a Hidden Markov Models Voice Activity Detector which a feature vector containing two components: normalized energy and delta energy. The key points of this Thesis are the following: 1) feature vector extension providing spectral information, 2) Hidden Markov Models adjustment to environment and study of different Hidden Markov Model topologies and, finally, 3) study and inclusion of new features, different from point 1, to reject the pronunciations coming from far-field speakers. Detection results, taking into account the above three points, show the advantages of using this method and are significantly better than the results obtained under the same conditions by other well-known voice activity detectors.

Generating segmental foreign accent

For most of us, speaking in a non-native language involves deviating to some extent from native pronunciation norms. However, the detailed basis for foreign accent (FA) remains elusive, in part due to methodological challenges in isolating segmental from suprasegmental factors. The current study examines the role of segmental features in conveying FA through the use of a generative approach in which accent is localised to single consonantal segments. Three techniques are evaluated: the first requires a highly-proficiency bilingual to produce words with isolated accented segments; the second uses cross-splicing of context-dependent consonants from the non-native language into native words; the third employs hidden Markov model synthesis to blend voice models for both languages. Using English and Spanish as the native/non-native languages respectively, listener cohorts from both languages identified words and rated their degree of FA. All techniques were capable of generating accented words, but to differing degrees. Naturally-produced speech led to the strongest FA ratings and synthetic speech the weakest, which we interpret as the outcome of over-smoothing. Nevertheless, the flexibility offered by synthesising localised accent encourages further development of the method.

Human activity recognition by inertial signals obtained from a smartphone

Human Activity Recognition (HAR) is an emerging research field with the aim to identify the actions carried out by a person given a set of observations and the surrounding environment. The wide growth in this research field inside the scientific community is mainly explained by the high number of applications that are arising in the last years. A great part of the most promising applications are related to the healthcare field, where it is possible to track the mobility of patients with motor dysfunction as also the physical activity in patients with cardiovascular risk. Until a few years ago, by using distinct kind of sensors, a patient follow-up was possible. However, far from being a long-term solution and with the smartphone irruption, that monitoring can be achieved in a non-invasive way by using the embedded smartphone’s sensors. For these reasons this Final Degree Project arises with the main target to evaluate new feature extraction techniques in order to carry out an activity and user recognition, and also an activity segmentation. The recognition is done thanks to the inertial signals integration obtained by two widespread sensors in the greater part of smartphones: accelerometer and gyroscope. In particular, six different activities are evaluated walking, walking-upstairs, walking-downstairs, sitting, standing and lying. Furthermore, a segmentation task is carried out taking into account the activities performed by thirty users. This can be done by using Hidden Markov Models and also a set of tools tested satisfactory in speech recognition: HTK (Hidden Markov Model Toolkit).

Human activity recognition by inertial signals obtained from a smartphone

El Reconocimiento de Actividades Humanas es un área de investigación emergente, cuyo objetivo principal es identificar las acciones realizadas por un sujeto analizando las señales obtenidas a partir de unos sensores. El rápido crecimiento de este área de investigación dentro de la comunidad científica se explica, en parte, por el elevado número de aplicaciones que están surgiendo en los últimos años. Gran parte de las aplicaciones más prometedoras se encuentran en el campo de la salud, donde se puede hacer un seguimiento del nivel de movilidad de pacientes con trastornos motores, así como monitorizar el nivel de actividad física en pacientes con riesgo cardiovascular. Hasta hace unos años, mediante el uso de distintos tipos de sensores se podía hacer un seguimiento del paciente. Sin embargo, lejos de ser una solución a largo plazo y gracias a la irrupción del teléfono inteligente, este seguimiento se puede hacer de una manera menos invasiva, haciendo uso de la gran variedad de sensores integrados en este tipo de dispositivos. En este contexto nace este Trabajo de Fin de Grado, cuyo principal objetivo es evaluar nuevas técnicas de extracción de características para llevar a cabo un reconocimiento de actividades y usuarios así como una segmentación de aquellas. Este reconocimiento se hace posible mediante la integración de señales inerciales obtenidas por dos sensores presentes en la gran mayoría de teléfonos inteligentes: acelerómetro y giróscopo. Concretamente, se evalúan seis tipos de actividades realizadas por treinta usuarios: andar, subir escaleras, bajar escaleras, estar sentado, estar de pie y estar tumbado. Además y de forma paralela, se realiza una segmentación temporal de los distintos tipos de actividades realizadas por dichos usuarios. Todo ello se llevará a cabo haciendo uso de los Modelos Ocultos de Markov, así como de un conjunto de herramientas probadas satisfactoriamente en reconocimiento del habla: HTK (Hidden Markov Model Toolkit).

Regularized model learning in EDAs for continuous and multi-objective optimization

Probabilistic modeling is the de�ning characteristic of estimation of distribution algorithms (EDAs) which determines their behavior and performance in optimization. Regularization is a well-known statistical technique used for obtaining an improved model by reducing the generalization error of estimation, especially in high-dimensional problems. `1-regularization is a type of this technique with the appealing variable selection property which results in sparse model estimations. In this thesis, we study the use of regularization techniques for model learning in EDAs. Several methods for regularized model estimation in continuous domains based on a Gaussian distribution assumption are presented, and analyzed from di�erent aspects when used for optimization in a high-dimensional setting, where the population size of EDA has a logarithmic scale with respect to the number of variables. The optimization results obtained for a number of continuous problems with an increasing number of variables show that the proposed EDA based on regularized model estimation performs a more robust optimization, and is able to achieve signi�cantly better results for larger dimensions than other Gaussian-based EDAs. We also propose a method for learning a marginally factorized Gaussian Markov random �eld model using regularization techniques and a clustering algorithm. The experimental results show notable optimization performance on continuous additively decomposable problems when using this model estimation method. Our study also covers multi-objective optimization and we propose joint probabilistic modeling of variables and objectives in EDAs based on Bayesian networks, speci�cally models inspired from multi-dimensional Bayesian network classi�ers. It is shown that with this approach to modeling, two new types of relationships are encoded in the estimated models in addition to the variable relationships captured in other EDAs: objectivevariable and objective-objective relationships. An extensive experimental study shows the e�ectiveness of this approach for multi- and many-objective optimization. With the proposed joint variable-objective modeling, in addition to the Pareto set approximation, the algorithm is also able to obtain an estimation of the multi-objective problem structure. Finally, the study of multi-objective optimization based on joint probabilistic modeling is extended to noisy domains, where the noise in objective values is represented by intervals. A new version of the Pareto dominance relation for ordering the solutions in these problems, namely �-degree Pareto dominance, is introduced and its properties are analyzed. We show that the ranking methods based on this dominance relation can result in competitive performance of EDAs with respect to the quality of the approximated Pareto sets. This dominance relation is then used together with a method for joint probabilistic modeling based on `1-regularization for multi-objective feature subset selection in classi�cation, where six di�erent measures of accuracy are considered as objectives with interval values. The individual assessment of the proposed joint probabilistic modeling and solution ranking methods on datasets with small-medium dimensionality, when using two di�erent Bayesian classi�ers, shows that comparable or better Pareto sets of feature subsets are approximated in comparison to standard methods.

HF spectrum activity prediction model based on HMM for cognitive radio applications

Although most of the research on Cognitive Radio is focused on communication bands above the HF upper limit (30 MHz), Cognitive Radio principles can also be applied to HF communications to make use of the extremely scarce spectrum more efficiently. In this work we consider legacy users as primary users since these users transmit without resorting to any smart procedure, and our stations using the HFDVL (HF Data+Voice Link) architecture as secondary users. Our goal is to enhance an efficient use of the HF band by detecting the presence of uncoordinated primary users and avoiding collisions with them while transmitting in different HF channels using our broad-band HF transceiver. A model of the primary user activity dynamics in the HF band is developed in this work to make short-term predictions of the sojourn time of a primary user in the band and avoid collisions. It is based on Hidden Markov Models (HMM) which are a powerful tool for modelling stochastic random processes and are trained with real measurements of the 14 MHz band. By using the proposed HMM based model, the prediction model achieves an average 10.3% prediction error rate with one minute-long channel knowledge but it can be reduced when this knowledge is extended: with the previous 8 min knowledge, an average 5.8% prediction error rate is achieved. These results suggest that the resulting activity model for the HF band could actually be used to predict primary users activity and included in a future HF cognitive radio based station.

Cooperative off-policy prediction of markov decision processes in adaptive networks

We apply diffusion strategies to propose a cooperative reinforcement learning algorithm, in which agents in a network communicate with their neighbors to improve predictions about their environment. The algorithm is suitable to learn off-policy even in large state spaces. We provide a mean-square-error performance analysis under constant step-sizes. The gain of cooperation in the form of more stability and less bias and variance in the prediction error, is illustrated in the context of a classical model. We show that the improvement in performance is especially significant when the behavior policy of the agents is different from the target policy under evaluation.

On-road visual vehicle tracking using Markov chain Monte Carlo with metropolis sampling

In this study, a method for vehicle tracking through video analysis based on Markov chain Monte Carlo (MCMC) particle filtering with metropolis sampling is proposed. The method handles multiple targets with low computational requirements and is, therefore, ideally suited for advanced-driver assistance systems that involve real-time operation. The method exploits the removed perspective domain given by inverse perspective mapping (IPM) to define a fast and efficient likelihood model. Additionally, the method encompasses an interaction model using Markov Random Fields (MRF) that allows treatment of dependencies between the motions of targets. The proposed method is tested in highway sequences and compared to state-of-the-art methods for vehicle tracking, i.e., independent target tracking with Kalman filtering (KF) and joint tracking with particle filtering. The results showed fewer tracking failures using the proposed method.