Speech is by far the most natural form of communication between human beings. It is intuitive, expressive and contains information at several cognitive levels. We as humans, are perceptive to several of these cognitive levels of information, as we can gather the information pertaining to the identity of the speaker, the speaker's gender, emotion, location, the language, and so on, in addition to the content of what is being spoken. This makes speech based human machine interaction (HMI), both desirable and challenging for the same set of reasons. For HMI to be natural for humans, it is imperative that a machine understands information present in speech, at least at the level of speaker identity, language, location in space, and the summary of what is being spoken. Although one can draw parallels between the human-human interaction and HMI, the two differ in their purpose. We, as humans, interact with a machine, mostly in the context of getting a task done more efficiently, than is possible without the machine. Thus, typically in HMI, controlling the machine in a specific manner is the primary goal. In this context, it can be argued that, HMI, with a limited vocabulary containing specific commands, would suffice for a more efficient use of the machine. In this thesis, we address the problem of ``Who spoke what and where", in the context of a machine understanding the information pertaining to identities of the speakers, their locations in space and the keywords they spoke, thus considering three levels of information - speaker identity (who), location (where) and keywords (what). This can be addressed with the help of multiple sensors like microphones, video camera, proximity sensors, motion detectors, etc., and combining all these modalities. However, we explore the use of only microphones to address this issue. In practical scenarios, often there are times, wherein, multiple people are talking at the same time. Thus, the goal of this thesis is to detect all the speakers, their keywords, and their locations in mixture signals containing speech from simultaneous speakers. Addressing this problem of ``Who spoke what and where" using only microphone signals, forms a part of acoustic scene analysis (ASA) of speech based acoustic events. We divide the problem of ``who spoke what and where" into two sub-problems: ``Who spoke what?" and ``Who spoke where". Each of these problems is cast in a generic latent variable (LV) framework to capture information in speech at different levels. We associate a LV to represent each of these levels and model the relationship between the levels using conditional dependency. The sub-problem of ``who spoke what" is addressed using single channel microphone signal, by modeling the mixture signal in terms of LV mass functions of speaker identity, the conditional mass function of the keyword spoken given the speaker identity, and a speaker-specific-keyword model. The LV mass functions are estimated in a Maximum likelihood (ML) framework using the Expectation Maximization (EM) algorithm using Student's-t Mixture Model (tMM) as speaker-specific-keyword models. Motivated by HMI in a home environment, we have created our own database. In mixture signals, containing two speakers uttering the keywords simultaneously, the proposed framework achieves an accuracy of 82 % for detecting both the speakers and their respective keywords. The other sub-problem of ``who spoke where?" is addressed in two stages. In the first stage, the enclosure is discretized into sectors. The speakers and the sectors in which they are located are detected in an approach similar to the one employed for ``who spoke what" using signals collected from a Uniform Circular Array (UCA). However, in place of speaker-specific-keyword models, we use tMM based speaker models trained on clean speech, along with a simple Delay and Sum Beamformer (DSB). In the second stage, the speakers are localized within the active sectors using a novel region constrained localization technique based on time difference of arrival (TDOA). Since the problem being addressed is a multi-label classification task, we use the average Hamming score (accuracy) as the performance metric. Although the proposed approach yields an accuracy of 100 % in an anechoic setting for detecting both the speakers and their corresponding sectors in two-speaker mixture signals, the performance degrades to an accuracy of 67 % in a reverberant setting, with a $60$ dB reverberation time (RT60) of 300 ms. To improve the performance under reverberation, prior knowledge of the location of multiple sources is derived using a novel technique derived from geometrical insights into TDOA estimation. With this prior knowledge, the accuracy of the proposed approach improves to 91 %. It is worthwhile to note that, the accuracies are computed for mixture signals containing more than 90 % overlap of competing speakers. The proposed LV framework offers a convenient methodology to represent information at broad levels. In this thesis, we have shown its use with three different levels. This can be extended to several such levels to be applicable for a generic analysis of the acoustic scene consisting of broad levels of events. It will turn out that not all levels are dependent on each other and hence the LV dependencies can be minimized by independence assumption, which will lead to solving several smaller sub-problems, as we have shown above. The LV framework is also attractive to incorporate prior knowledge about the acoustic setting, which is combined with the evidence from the data to derive the information about the presence of an acoustic event. The performance of the framework, is dependent on the choice of stochastic models, which model the likelihood function of the data given the presence of acoustic events. However, it provides an access to compare and contrast the use of different stochastic models for representing the likelihood function.

Ce travail de thèse poursuit une perspective double dans l'usage conjoint des méthodes de Monte Carlo séquentielles (MMS) et de l'algorithme Espérance-Maximisation (EM) dans le cadre des modèles de Markov cachés présentant une structure de dépendance markovienne d'ordre supérieur à 1 au niveau de la composante inobservée. Tout d'abord, nous commençons par un exposé succinct de l'assise théorique des deux concepts statistiques à Travers les chapitres 1 et 2 qui leurs sont consacrés. Dans un second temps, nous nous intéressons à la mise en pratique simultanée des deux concepts au chapitre 3 et ce dans le cadre usuel ou la structure de dépendance est d'ordre 1, l'apport des méthodes MMS dans ce travail réside dans leur capacité à approximer efficacement des fonctionnelles conditionnelles bornées, notamment des quantités de filtrage et de lissage dans un cadre non linéaire et non gaussien. Quant à l'algorithme EM, il est motivé par la présence à la fois de variables observables, et inobservables (ou partiellement observées) dans les modèles de Markov Cachés et singulièrement les modèles de volatilité stochastique étudié. Après avoir présenté aussi bien l'algorithme EM que les méthodes MCS ainsi que quelques une de leurs propriétés dans les chapitres 1 et 2 respectivement, nous illustrons ces deux outils statistiques au travers de la calibration d'un modèle de volatilité stochastique. Cette application est effectuée pour des taux change ainsi que pour quelques indices boursiers au chapitre 3. Nous concluons ce chapitre sur un léger écart du modèle de volatilité stochastique canonique utilisé ainsi que des simulations de Monte Carlo portant sur le modèle résultant. Enfin, nous nous efforçons dans les chapitres 4 et 5 à fournir les assises théoriques et pratiques de l'extension des méthodes Monte Carlo séquentielles notamment le filtrage et le lissage particulaire lorsque la structure markovienne est plus prononcée. En guise d’illustration, nous donnons l'exemple d'un modèle de volatilité stochastique dégénéré dont une approximation présente une telle propriété de dépendance. / This thesis pursues a double perspective in the joint use of sequential Monte Carlo methods (SMC) and the Expectation-Maximization algorithm (EM) under hidden Mar­kov models having a Markov dependence structure of order grater than one in the unobserved component signal. Firstly, we begin with a brief description of the theo­retical basis of both statistical concepts through Chapters 1 and 2 that are devoted. In a second hand, we focus on the simultaneous implementation of both concepts in Chapter 3 in the usual setting where the dependence structure is of order 1. The contribution of SMC methods in this work lies in their ability to effectively approximate any bounded conditional functional in particular, those of filtering and smoothing quantities in a non-linear and non-Gaussian settings. The EM algorithm is itself motivated by the presence of both observable and unobservable ( or partially observed) variables in Hidden Markov Models and particularly the stochastic volatility models in study. Having presented the EM algorithm as well as the SMC methods and some of their properties in Chapters 1 and 2 respectively, we illustrate these two statistical tools through the calibration of a stochastic volatility model. This application is clone for exchange rates and for some stock indexes in Chapter 3. We conclude this chapter on a slight departure from canonical stochastic volatility model as well Monte Carlo simulations on the resulting model. Finally, we strive in Chapters 4 and 5 to provide the theoretical and practical foundation of sequential Monte Carlo methods extension including particle filtering and smoothing when the Markov structure is more pronounced. As an illustration, we give the example of a degenerate stochastic volatility model whose approximation has such a dependence property.

Les batteries électriques sont omniprésentes dans notre vie quotidienne : ordinateur, téléphone, etc. Elles jouent un rôle important dans le défi de la transition énergétique : anticiper la raréfaction des énergies fossiles et réduire la pollution, en développant le stockage des énergies renouvelables et les transports électriques. Cependant, l'estimation de l'état de charge (State of Charge – SoC) d'une batterie est difficile et les modèles de prédiction actuels sont peu robustes. En effet, une batterie est un système électrochimique complexe, dont la dynamique est influencée non seulement par ses caractéristiques internes, mais aussi par les conditions d'usages souvent non contrôlables : température, profil d’utilisation, etc. Or, une estimation précise du SoC permet de garantir une utilisation sûre de la batterie en évitant une surcharge ou surdécharge ; mais aussi d’estimer son autonomie. Dans cette étude, nous utilisons un modèle à espaces d'états gouverné par une chaîne de Markov cachée. Ce modèle est fondé sur des équations physiques et la chaîne de Markov cachée permet d’appréhender les différents «régimes de fonctionnement» de la batterie. Pour garantir l’unicité des paramètres du modèle, nous démontrons son identifiabilité à partir de contraintes simples et naturelles sur ses paramètres «physiques ». L’estimation du SoC dans un véhicule électrique doit être faîte en ligne et avec une puissance de calcul limitée. Nous estimons donc le SoC en utilisant une technique d’échantillonnage préférentiel séquentiel. D’autre part l’estimation des paramètres est faîte à partir d’une base d’apprentissage pour laquelle les états de la chaîne de Markov et le SoC ne sont pas observés. Nous développons et testons trois algorithmes adaptés à notre modèle à structure latente : un échantillonneur particulaire de Gibbs, un algorithme de Monte-Carlo EM pénalisé par des contraintes d’identifiabilité et un algorithme de Monte-Carlo EM pénalisé par une loi a priori. Par ailleurs les états cachés de la chaîne de Markov visent à modéliser les différents régimes du fonctionnement de la batterie. Nous identifions leur nombre par divers critères de sélection de modèles. Enfin, à partir de données issues de trois types de batteries (cellule, module et pack d’un véhicule électrique), notre modèle a permis d’appréhender les différentes sollicitations de la batterie et donne des estimations robustes et précises du SoC. / Electric batteries are omnipresent in our daily lives: computers, smartphones, etc. Batteries are important for anticipating the scarcity of fossil fuels and tackling their environmental impact. Therefore, estimating the State of Charge (SoC) of a battery is nowadays a challenging issue, as existing physical and statistical models are not yet robust. Indeed a battery is a complex electrochemical system. Its dynamic depends not only on its internal characteristics but also on uncontrolled usage conditions: temperature, usage profile, etc. However the SoC estimation helps to prevent overcharge and deep discharge, and to estimate the battery autonomy. In this study, the battery dynamics are described by a set of physical linear equations, switching randomly according to a Markov chain. This model is referred to as switching Markov state space model. To ensure the unicity of the model parameters, we prove its identifiability by applying straightforward and natural constraints on its “physical” parameters. Embedded applications, like electric vehicles, impose online estimated with hardware and time constraints. Therefore we estimate the SoC using a sequential importance sampling technique. Furthermore the model includes two latent variables: the SoC and the Markov chain state. Thus, to estimate the parameters, we develop and test three algorithms adapted to latent structure models: particle Gibbs sampler, Monte Carlo EM penalized with identifiability constraints, and Monte Carlo EM penalized with a prior distribution. The hidden Markov states aim to model the different “regimes” of the battery dynamics. We identify their number using different model selection criteria. Finally, when applied to various data from three battery types (cell, module and pack of an electric vehicle) our model allows us to analyze the battery dynamics and to obtain a robust and accurate SoC estimation under uncontrolled usage conditions.

This thesis deals with state estimation methods for AC drives sensorless control and with possibilities of the estimation. Conditions for observability for a synchronous drive were derived, as well as conditions for the moment of inertia and the load torque observability for both drive types - synchronous and asynchronous. The possibilities of the estimation were confirmed by experimental results. The covariance matrices for all filters were found using an EM algorithm. Both drives were also identified. The algoritms used for state estimation are Extended Kalman Filter, Unscented Kalman Filter, Particle Filters and Moving Horizon Estimator.

This master's thesis deals with with FIT-Miner, the system for knowledge discovery in databases. The first part of this paper describes the data-mining process, mixture model's issues and FIT-Miner system. Second part deals with design, implementation and testing of created module, which is used for cluster analysis with Expectation-Maximalization algorithm. The end of the paper is focused to design of modules using Java Store Procedures Technology.

Zapisovač letových údajů je zařízení určené pro zaznamenávání letových dat z různých senzorů v letadlech. Analýza letových údajů hraje důležitou roli ve vývoji a testování avioniky. Testování a hodnocení charakteristik letadla se často provádí pomocí testovacích manévrů. Naměřená data z jednoho letu jsou uložena v jednom letovém záznamu, který může obsahovat několik testovacích manévrů. Cílem této práce je identi kovat základní testovací manévry s pomocí naměřených letových dat. Teoretická část popisuje letové manévry a formát měřených letových dat. Analytická část popisuje výzkum v oblasti klasi kace založené na statistice a teorii pravděpodobnosti potřebnou pro pochopení složitých Gaussovských směšovacích modelů. Práce uvádí implementaci, kde jsou Gaussovy směšovací modely použité pro klasifi kaci testovacích manévrů. Navržené řešení bylo testováno pro data získána z letového simulátoru a ze skutečného letadla. Ukázalo se, že Gaussovy směšovací modely poskytují vhodné řešení pro tento úkol. Další možný vývoj práce je popsán v závěrečné kapitole.

In Natural Language Processing (NLP), speech and text are parsed and generated with language models and parser models, and translated with translation models. Each model contains a set of numerical parameters which are found by applying a suitable training algorithm to a set of training data. Many such training algorithms are instances of the Expectation-Maximization (EM) algorithm. In [BSV15], a generic EM algorithm for NLP is described. This work presents a particular speech model, the Hidden Markov model, and its standard training algorithm, the Baum-Welch algorithm. It is then shown that the Baum-Welch algorithm is an instance of the generic EM algorithm introduced by [BSV15], from which follows that all statements about the generic EM algorithm also apply to the Baum-Welch algorithm, especially its correctness and convergence properties.:1 Introduction 1.1 N-gram models 1.2 Hidden Markov model 2 Expectation-maximization algorithms 2.1 Preliminaries 2.2 Algorithmic skeleton 2.3 Corpus-based step mapping 2.4 Simple counting step mapping 2.5 Regular tree grammars 2.6 Inside-outside step mapping 2.7 Review 3 The Hidden Markov model 3.1 Forward and backward algorithms 3.2 The Baum-Welch algorithm 3.3 Deriving the Baum-Welch algorithm 3.3.1 Model parameter and countable events 3.3.2 Tree-shaped hidden information 3.3.3 Complete-data corpus 3.3.4 Inside weights 3.3.5 Outside weights 3.3.6 Complete-data corpus (cont.) 3.3.7 Step mapping 3.4 Review Appendix A Elided proofs from Chapter 3 A.1 Proof of Lemma 3.8 A.2 Proof of Lemma 3.9 B Formulary for Chapter 3 Bibliography

In this thesis, we develop some inferential results for one-shot device testing data analysis. These extend and generalize existing methods in the literature. First, a competing-risk model is introduced for one-shot testing data under accelerated life-tests. One-shot devices are products which will be destroyed immediately after use. Therefore, we can observe only a binary status as data, success or failure, of such products instead of its lifetime. Many one-shot devices contain multiple components and failure of any one of them will lead to the failure of the device. Failed devices are inspected to identify the specific cause of failure. Since the exact lifetime is not observed, EM algorithm becomes a natural tool to obtain the maximum likelihood estimates of the model parameters. Here, we develop the EM algorithm for competing exponential and Weibull cases. Second, a semi-parametric approach is developed for simple one-shot device testing data. Semi-parametric estimation is a model that consists of parametric and non-parametric components. For this purpose, we only assume the hazards at different stress levels are proportional to each other, but no distributional assumption is made on the lifetimes. This provides a greater flexibility in model fitting and enables us to examine the relationship between the reliability of devices and the stress factors. Third, Bayesian inference is developed for one-shot device testing data under exponential distribution and Weibull distribution with non-constant shape parameters for competing risks. Bayesian framework provides statistical inference from another perspective. It assumes the model parameters to be random and then improves the inference by incorporating expert's experience as prior information. This method is shown to be very useful if we have limited failure observation wherein the maximum likelihood estimator may not exist. The thesis proceeds as follows. In Chapter 2, we assume the one-shot devices to have two components with lifetimes having exponential distributions with multiple stress factors. We then develop an EM algorithm for developing likelihood inference for the model parameters as well as some useful reliability characteristics. In Chapter 3, we generalize to the situation when lifetimes follow a Weibull distribution with non-constant shape parameters. In Chapter 4, we propose a semi-parametric model for simple one-shot device test data based on proportional hazards model and develop associated inferential results. In Chapter 5, we consider the competing risk model with exponential lifetimes and develop inference by adopting the Bayesian approach. In Chapter 6, we generalize these results on Bayesian inference to the situation when the lifetimes have a Weibull distribution. Finally, we provide some concluding remarks and indicate some future research directions in Chapter 7. / Thesis / Doctor of Philosophy (PhD)

No description available.

Le modèle GARCH à changement de régimes est le fondement de cette thèse. Ce modèle offre de riches dynamiques pour modéliser les données financières en combinant une structure GARCH avec des paramètres qui varient dans le temps. Cette flexibilité donne malheureusement lieu à un problème de path dependence, qui a empêché l'estimation du modèle par le maximum de vraisemblance depuis son introduction, il y a déjà près de 20 ans. La première moitié de cette thèse procure une solution à ce problème en développant deux méthodologies permettant de calculer l'estimateur du maximum de vraisemblance du modèle GARCH à changement de régimes. La première technique d'estimation proposée est basée sur l'algorithme Monte Carlo EM et sur l'échantillonnage préférentiel, tandis que la deuxième consiste en la généralisation des approximations du modèle introduites dans les deux dernières décennies, connues sous le nom de collapsing procedures. Cette généralisation permet d'établir un lien méthodologique entre ces approximations et le filtre particulaire. La découverte de cette relation est importante, car elle permet de justifier la validité de l'approche dite par collapsing pour estimer le modèle GARCH à changement de régimes. La deuxième moitié de cette thèse tire sa motivation de la crise financière de la fin des années 2000 pendant laquelle une mauvaise évaluation des risques au sein de plusieurs compagnies financières a entraîné de nombreux échecs institutionnels. À l'aide d'un large éventail de 78 modèles économétriques, dont plusieurs généralisations du modèle GARCH à changement de régimes, il est démontré que le risque de modèle joue un rôle très important dans l'évaluation et la gestion du risque d'investissement à long terme dans le cadre des fonds distincts. Bien que la littérature financière a dévoué beaucoup de recherche pour faire progresser les modèles économétriques dans le but d'améliorer la tarification et la couverture des produits financiers, les approches permettant de mesurer l'efficacité d'une stratégie de couverture dynamique ont peu évolué. Cette thèse offre une contribution méthodologique dans ce domaine en proposant un cadre statistique, basé sur la régression, permettant de mieux mesurer cette efficacité. / The Markov-switching GARCH model is the foundation of this thesis. This model offers rich dynamics to model financial data by allowing for a GARCH structure with time-varying parameters. This flexibility is unfortunately undermined by a path dependence problem which has prevented maximum likelihood estimation of this model since its introduction, almost 20 years ago. The first half of this thesis provides a solution to this problem by developing two original estimation approaches allowing us to calculate the maximum likelihood estimator of the Markov-switching GARCH model. The first method is based on both the Monte Carlo expectation-maximization algorithm and importance sampling, while the second consists of a generalization of previously proposed approximations of the model, known as collapsing procedures. This generalization establishes a novel relationship in the econometric literature between particle filtering and collapsing procedures. The discovery of this relationship is important because it provides the missing link needed to justify the validity of the collapsing approach for estimating the Markov-switching GARCH model. The second half of this thesis is motivated by the events of the financial crisis of the late 2000s during which numerous institutional failures occurred because risk exposures were inappropriately measured. Using 78 different econometric models, including many generalizations of the Markov-switching GARCH model, it is shown that model risk plays an important role in the measurement and management of long-term investment risk in the context of variable annuities. Although the finance literature has devoted a lot of research into the development of advanced models for improving pricing and hedging performance, the approaches for measuring dynamic hedging effectiveness have evolved little. This thesis offers a methodological contribution in this area by proposing a statistical framework, based on regression analysis, for measuring the effectiveness of dynamic hedges for long-term investment guarantees.