This dissertation focuses on improving content-based recommendation systems for music. Specifically, progress in the development in music content-based recommendation systems has stalled in recent years due to some faulty assumptions: 1. most acoustic content-based systems for music information retrieval (MIR) assume a bag-of-frames model, where it is assumed that a song contains a simplistic, global audio texture 2. genre, style, mood, and authors are appropriate categories for machine-oriented recommendation 3. similarity is a universal construct and does not vary among different users The main contribution of this dissertation is to address these faulty assumptions by describing a novel approach in MIR that provides user-centric, content-based recommendations based on statistics of acoustic sound elements. First, this dissertation presents the acoustic segment modeling framework that describes a piece of music as a temporal sequence of acoustic segment models (ASMs), which represent individual polyphonic sound elements. A dictionary of ASMs generated in an unsupervised process defines a vocabulary of acoustic tokens that are able to transcribe new musical pieces. Next, standard text-based information retrieval algorithms use statistics of ASM counts to perform various retrieval tasks. Despite a simple feature set compared to other content-based genre recommendation algorithms, the acoustic segment modeling approach is highly competitive on standard genre classification databases. Fundamental to the success of the acoustic segment modeling approach is the ability to model acoustical semantics in a musical piece, which is demonstrated by the detection of musical attributes on temporal characteristics. Further, it is shown that the acoustic segment modeling procedure is able to capture the inherent structure of melody by providing near state-of-the-art performance on an automatic chord recognition task. This dissertation demonstrates that some classification tasks, such as genre, possess information that is not contained in the acoustic signal; therefore, attempts at modeling these categories using only the acoustic content is ill-fated. Further, notions of music similarity are personal in nature and are not derived from a universal ontology. Therefore, this dissertation addresses the second and third limitation of previous content-based retrieval approaches by presenting a user-centric preference rating algorithm. Individual users possess their own cognitive construct of similarity; therefore, retrieval algorithms must demonstrate this flexibility. The proposed rating algorithm is based on the principle of minimum classification error (MCE) training, which has been demonstrated to be robust against outliers and also minimizes the Parzen estimate of the theoretical classification risk. The outlier immunity property limits the effect of labels that arise from non-content-based sources. The MCE-based algorithm performs better than a similar ratings prediction algorithm. Further, this dissertation discusses extensions and future work.

Acoustic modeling in state-of-the-art speech recognition systems is commonly based on discriminative criteria. Different from the paradigm of the conventional distribution estimation such as maximum a posteriori (MAP) and maximum likelihood (ML), the most popular discriminative criteria such as MCE and MPE aim at direct minimization of the empirical error rate. As recent ASR applications become diverse, it has been increasingly recognized that realistic applications often require a model that can be optimized for a task-specific goal or a particular scenario beyond the general purposes of the current discriminative criteria. These specific requirements cannot be directly handled by the current discriminative criteria since the objective of the criteria is to minimize the overall empirical error rate. In this thesis, we propose novel objective-driven discriminative training and adaptation frameworks, which are generalized from the minimum classification error (MCE) criterion, for various tasks and scenarios of speech recognition and detection. The proposed frameworks are constructed to formulate new discriminative criteria which satisfy various requirements of the recent ASR applications. In this thesis, each objective required by an application or a developer is directly embedded into the learning criterion. Then, the objective-driven discriminative criterion is used to optimize an acoustic model in order to achieve the required objective. Three task-specific requirements that the recent ASR applications often require in practice are mainly taken into account in developing the objective-driven discriminative criteria. First, an issue of individual error minimization of speech recognition is addressed and we propose a direct minimization algorithm for each error type of speech recognition. Second, a rapid adaptation scenario is embedded into formulating discriminative linear transforms under the MCE criterion. A regularized MCE criterion is proposed to efficiently improve the generalization capability of the MCE estimate in a rapid adaptation scenario. Finally, the particular operating scenario that requires a system model optimized at a given specific operating point is discussed over the conventional receiver operating characteristic (ROC) optimization. A constrained discriminative training algorithm which can directly optimize a system model for any particular operating need is proposed. For each of the developed algorithms, we provide an analytical solution and an appropriate optimization procedure.

Les travaux présentés dans cette thèse s’appliquent à la caractérisation de propriétés mécaniques par la microscopie acoustique. Ils décrivent un capteur focalisé innovant qui autorise à la fois une topographie et une imagerie quantitative d’un matériau élastique. L’innovation consiste en la séparation des différents modes de propagation d’un matériau excité par une sonde focalisée multiélément. La mesure par temps de vol de la vitesse de propagation des modes de surfaces de matériaux élastiques et anisotropes offre une possibilité de quantification du module caractérisant l’élasticité : le module de Young. Le dimensionnement de la sonde multiélément qui est décrit ici est rendu possible grâce au développement d’un modèle de champs acoustiques permettant d’anticiper le champ rayonné par chaque élément. Un deuxième modèle traitant de l’étude temporel des signaux reçus par la sonde focalisée est aussi présenté pour vérifier le comportement discriminant de la sonde des différentes ondes pouvant se propager. La mesure de propriétés mécaniques par la sonde focalisée est appliquée à différents échantillons et propose des résultats cohérents avec une grande sensibilité. La possibilité de réaliser des images de propriétés mécaniques est ainsi démontrée. D’abord adaptée pour des fréquences de l’ordre de la trentaine de mégahertz, cette sonde possède un nombre limité d’éléments pour assurer une simplicité de conception et de fabrication permettant par la suite une miniaturisation du capteur pour atteindre des fréquences proches du gigahertz. / The work presented in this thesis is applied to the characterization of mechanical properties by acoustic microscopy. It describes an innovative focused sensor that enables both topography and quantitative imaging of an elastic material. The innovation consists in the separation of the different propagation modes of a material excited by a focused multielement probe. Measuring the surface mode propagation velocity of elastic and anisotropic materials thanks to their time of flight provides a possibility of quantifying the module characterizing the elasticity: the Young's modulus. The dimensions of the multielement probe are described here and rely on an acoustic field model developed to anticipate the field radiated by each element. A second model studies the temporal behaviour of the focused probe and also verifies the discrimination of the different waves that propagate. The measurement of mechanical properties by the multielement probe is applied to different samples and provides consistent results with high sensitivity. The ability to produce images of mechanical properties is thus demonstrated. First suitable for frequencies near thirty megahertz, this sensor has a limited number of elements to ensure a simplicity of design and manufacture for a subsequent miniaturization of the sensor to achieve frequencies near the gigahertz.

L’objectif de cette thèse est de décrire de manière analytique le spectre du bruit de friction en mettant en évidence l’affiliation des pics spectraux aux cavités du conduit vocal, par le biais de la modélisation acoustique et avec l’aide des données IRM de 7 langues [30 locuteurs]. Les résultats sont les suivants : 1. La dispersion des sibilantes dans l’espace articulatoire dépend du système phonologique [contrastes [+/- antérieur], [+/- distribué], ou les deux]. En français [+/- antérieur], 7 locuteurs], la variation inter-individuelle est importante. 2. Cette variation est due à deux variantes articulatoires du /ʃ/ : (a) plutôt apical, comportant une cavité sublinguale, accompagné de protrusion labiale, et semblable au /ʂ/ polonais ; et [b] palatalisé, mettant en œuvre le bombement du dos de la langue, comparable au /ɕ/ polonais. L’équivalence acoustique des deux variantes est démontrée par une simulation acoustique systématique. 3. En polonais, où la différence articulatoire est phonémique, /ʂ/ est caractérisé par un pic ultra-bas [1,5-1,8 kHz], affilié à la cavité antérieure, d’après la simulation acoustique à l’aide de fonction d’aire réelles de deux locuteurs. 4. Les données articulatoires présentent systématiquement une constriction dentale étroite. D’après la modélisation acoustique, la protrusion labiale aurait comme effet d’abaisser la fréquence d’un formant affilié spécifiquement à la cavité labiale. En somme, la présence de deux constrictions étroites linguale et dentale rend possible le contrôle quasi indépendant d’au moins deux résonances. Cette spécificité garantit aux sibilantes un bruit distinctif qui permet d’expliquer la richesse de leurs inventaires / The aim of this study is to analytically describe the frication noise spectrum in terms of formant affiliation to vocal tract cavities. The high-resolution, teeth-inserted MRI data of sibilants in 7 languages [30 subjects in total] as well as 1D and 3D acoustic modeling are involved. The results are summarized as follows: 1. Sibilants’ dispersion within the articulatory space depends on the language’s phonemic inventory [with contrasts involving [+/- anterior], [+/- distributed], or both features]. A large amount of inter-speaker variation [7 subjects] is observed in French /s/ and /ʃ/ contrasted by [+/- anterior]]. 2. This variation is due to two articulatory variants of the French /ʃ/ : [a] apical, with a sublingual cavity, and protruded lips, like Polish /ʂ/ ; [b] palatalized, with a domed tongue dorsum, like Polish /ɕ/. Systematic acoustic modeling provides evidences about their acoustic equivalence. 3. In Polish, where /ʂ/ and /ɕ/ are contrastive, a super-low peak [1.5 – 1.8 kHz] characterizes the former. The acoustic modeling results using realistic area functions of two subjects show that this peak is affiliated to the front oral cavity. 4. The articulatory data show a systematic narrow constriction at the teeth in all of the examined sibilants. Acoustic modeling shows that lip protrusion results in lowering the formant affiliated specifically to the lip cavity. To conclude, the narrow tongue and teeth constrictions in sibilants allow the speaker to control quasi independently

Dans cette thèse, nous proposons d’utiliser des techniques fondées sur l’analyse factorielle pour la modélisation acoustique pour le traitement automatique de la parole, notamment pour la Reconnaissance Automatique de la parole. Nous nous sommes, dans un premier temps, intéressés à la réduction de l’empreinte mémoire des modèles acoustiques. Notre méthode à base d’analyse factorielle a démontré une capacité de mutualisation des paramètres des modèles acoustiques, tout en maintenant des performances similaires à celles des modèles de base. La modélisation proposée nous conduit à décomposer l’ensemble des paramètres des modèles acoustiques en sous-ensembles de paramètres indépendants, ce qui permet une grande flexibilité pour d’éventuelles adaptations (locuteurs, genre, nouvelles tâches).Dans les modélisations actuelles, un état d’un Modèle de Markov Caché (MMC) est représenté par un mélange de Gaussiennes (GMM : Gaussian Mixture Model). Nous proposons, comme alternative, une représentation vectorielle des états : les fac- teur d’états. Ces facteur d’états nous permettent de mesurer efficacement la similarité entre les états des MMC au moyen d’une distance euclidienne, par exemple. Grâce à cette représenation vectorielle, nous proposons une méthode simple et efficace pour la construction de modèles acoustiques avec des états partagés. Cette procédure s’avère encore plus efficace dans le cas de langues peu ou très peu dotées en ressouces et enconnaissances linguistiques. Enfin, nos efforts se sont portés sur la robustesse des systèmes de reconnaissance de la parole face aux variabilités acoustiques, et plus particulièrement celles générées par l’environnement. Nous nous sommes intéressés, dans nos différentes expérimentations, à la variabilité locuteur, à la variabilité canal et au bruit additif. Grâce à notre approche s’appuyant sur l’analyse factorielle, nous avons démontré la possibilité de modéliser ces différents types de variabilité acoustique nuisible comme une composante additive dans le domaine cepstral. Nous soustrayons cette composante des vecteurs cepstraux pour annuler son effet pénalisant pour la reconnaissance de la parole / In this thesis, we propose to use techniques based on factor analysis to build acoustic models for automatic speech processing, especially Automatic Speech Recognition (ASR). Frstly, we were interested in reducing the footprint memory of acoustic models. Our factor analysis-based method demonstrated that it is possible to pool the parameters of acoustic models and still maintain performance similar to the one obtained with the baseline models. The proposed modeling leads us to deconstruct the ensemble of the acoustic model parameters into independent parameter sub-sets, which allow a great flexibility for particular adaptations (speakers, genre, new tasks etc.). With current modeling techniques, the state of a Hidden Markov Model (HMM) is represented by a combination of Gaussians (GMM : Gaussian Mixture Model). We propose as an alternative a vector representation of states : the factors of states. These factors of states enable us to accurately measure the similarity between the states of the HMM by means of an euclidean distance for example. Using this vector represen- tation, we propose a simple and effective method for building acoustic models with shared states. This procedure is even more effective when applied to under-resourced languages. Finally, we concentrated our efforts on the robustness of the speech recognition sys- tems to acoustic variabilities, particularly those generated by the environment. In our various experiments, we examined speaker variability, channel variability and additive noise. Through our factor analysis-based approach, we demonstrated the possibility of modeling these different types of acoustic variability as an additive component in the cepstral domain. By compensation of this component from the cepstral vectors, we are able to cancel out the harmful effect it has on speech recognition

Nous nous intéressons au contrôle non destructif par ultrasons des matériaux industriels. En pratique, les signaux réceptionnés par le transducteur ultrasonore sont analysés pour détecter les discontinuités de la pièce inspectée. L'analyse est néanmoins rendue difficile par l'acquisition numérique, les effets de la propagation ultrasonore et la superposition des échos lorsque les discontinuités sont proches. La déconvolution parcimonieuse est une méthode inverse qui permet d'aborder ce problème afin de localiser précisément les discontinuités. Ce procédé favorise les signaux parcimonieux, c'est à dire ne contenant qu'un faible nombre de discontinuités. Dans la littérature, la déconvolution est généralement abordée sous l'hypothèse d'un modèle invariant en fonction de la distance de propagation, modalité qui n'est pas appropriée ici car l'onde se déforme au cours de son parcours et en fonction des discontinuités rencontrées. Cette thèse développe un modèle et des méthodes associées qui visent à annuler les dégradations dues à l'instrumentation et à la propagation ultrasonore, tout en résolvant des problèmes de superposition d'échos. Le premier axe consiste à modéliser la formation du signal ultrasonore en y intégrant les phénomènes propres aux ultrasons. Cette partie permet de construire un modèle linéaire mais non invariant, prenant en compte l'atténuation et la dispersion. L'étape de modélisation est validée par des acquisitions avec des matériaux atténuants. La deuxième partie de cette thèse concerne le développement de méthodes de déconvolution efficaces pour ce problème, reposant sur la minimisation d'un critère des moindres carrés pénalisé par la pseudo-norme L0. Nous avons développé des algorithmes d'optimisation spécifiques, prenant en compte, d'une part, un modèle de trains d'impulsions sur-échantillonné par rapport aux données, et d'autre part le caractère oscillant des formes d'onde ultrasonores. En utilisant des données synthétiques et expérimentales, ces algorithmes associés à un modèle direct adapté aboutissent à de meilleurs résultats comparés aux approches classiques pour un coût de calcul maîtrisé. Ces algorithmes sont finalement appliqués à des cas concrets de contrôle non destructif où ils démontrent leur efficacité. / This thesis deals with the ultrasonic non destructive testing of industrial parts. During real experiments, the signals received by the acoustic transducer are analyzed to detect the discontinuities of the part under test. This analysis can be a difficult task due to digital acquisition, propagation effects and echo overlapping if discontinuities are close. Sparse deconvolution is an inverse method that aims to estimate the precise positions of the discontinuities. The underlying hypothesis of this method is a sparse distribution of the solution, which means there are a few number of discontinuities. In the literature, deconvolution is addressed by a linear time-invariant model as a function of propagation distance, which in reality does not hold.The purpose of this thesis is therefore to develop a model and associated methods in order to cancel the effects of acquisition, propagation and echo overlapping. The first part is focused on the direct model development. In particular, we build a linear time-variant model that takes into account dispersive attenuation. This model is validated with experimental data acquired from attenuative materials. The second part of this work concerns the development of efficient sparse deconvolution algorithms, addressing the minimization of a least squares criterion penalized by a L0 pseudo-norm. Specific algorithms are developed for up-sampled deconvolution, and more robust exploration strategies are built for data containing oscillating waveforms. By using synthetic and experimental data, we show that the developed methods lead to better results compared to standard approaches for a competitive computation time. The proposed methods are then applied to real non destructive testing problems where they confirm their efficiency.

In the last decades, miniaturized flow chemistry has promised to bring the benefits of process intensification, continuous manufacturing and greener chemistry to the fine chemical industry. However, miniaturized catalytic processes where gas, liquid, and solids are involved have always been impeded by two main drawbacks: multiphase-flow maldistribution (i.e. gas channeling) and clogging of capillary reactors. In this thesis, first principle models have been used to capture the complexity of multiphase flow in micropacked-bed reactors, which can suffer from poor and unpredictable mass-transfer performance. When the particle size ranges 100 µm in diameter, capillary and viscous forces control the hydrodynamics. Under such conditions, the gas —and not the liquid— flows creating preferential channels that cause poor radial dispersion. Experimental observations from the literature were reproduced to validate a physical-based modeling approach, the Phase Field Method (PFM). This simulation strategy sheds light on the impact of the micropacked-bed geometry and wettability on the formation of preferential gas channels. Counterintuitively, to homogenize the two-phase flow hydrodynamics and reduce radial mass-transfer limitations, solvent wettability of the support needs to be restricted, showing best performance when the contact angle ranges 60° and capillary forces are still dominant. Visualization experiments showed that ultrasound irradiation can also be used to partially fluidized the bed and modify the hydrodynamics. Under sonication, residence time distributions (RTD) in micropacked-bed reactors revealed a two-order-of-magnitude reduction in dispersion, allowing for nearly plug-flow behavior at high gas and liquid flow rates. At a reduced scale, surfaces vibrating with a low amplitude were shown to fluidize, prevent and solve capillary tube blockage problems, which are commonly found in the fine chemical industry for continuous product synthesis. The modeling and simulation strategy used in this thesis, enables a fast prototyping methodology for the proper acoustic design of sonoreactors, whose scale-up was achieved by introducing slits in sonotrodes. In addition, a patent-pending helicoidal capillary sonoreactor has shown to transform longitudinal vibrating modes into radial and torsional modes, pioneering a new range of chemistry able to handle a high concentration of particles. The contributions of this thesis made in the fields of reaction engineering and process intensification have demonstrated how computational methods and experimental techniques in other areas of research can be used to foster innovation at a fast pace.

No description available.

No description available.

For an engine running at a constant speed, both exhaust and intake processes are periodic in nature. This inspires the muffler designer to go for the much easier and faster frequency domain modeling. But analogous to electrical filter, as per Thevenin’s theorem, the acoustic filter or muffler requires prior knowledge of the load-independent source characteristics (acoustic pressure and internal impedance), corresponding to the open circuit voltage and internal impedance of an electrical source. Studies have shown that it is not feasible to evaluate these source characteristics making use of either the direct measurement method or the indirect evaluation method. Hence, prediction of the radiated exhaust or intake noise has been subject to trial and error. Making use of the fact that pressure perturbation in a duct is a superposition of the forward moving wave and the reflected wave, a simple hybrid approach has been proposed making use of an interrelationship between progressive wave variables of the linear acoustic theory and Riemann variables of the method of characteristics. Neglecting the effect of nonlinearities, reflection of the forward moving wave has been duly incorporated at the exhaust valve. The reflection co-efficient of the system downstream of the exhaust valve has been calculated by means of the transfer matrix method at each of the several harmonics of the engine firing frequency. This simplified approach can predict exhaust noise with or without muffler for a naturally aspirated, single cylinder engine. However, this proves to be inadequate in predicting the exhaust noise of multi-cylinder engines. Thus, estimation of radiated noise has met only limited success in this approach. Strictly speaking, unique source characteristics do not exist for an IC engine because of the associated non-linearity of the time-varying source. Yet, a designer would like to know the un-muffled noise level in order to assess the required insertion loss of a suitable muffler. As far as the analysis and design of a muffler is concerned, the linear frequency-domain analysis by means of the transfer matrix approach is most convenient and time saving. Therefore, from a practical point of view, it is very desirable to be able to evaluate source characteristics, even if grossly approximate. If somehow it were possible to parameterize the source characteristics of an engine in terms of basic engine parameters, then it would be possible to evaluate the un-muffled noise before a design is taken up as a first approximation. This aspect has been investigated in detail in this work. A finite-volume CFD (one dimensional) model has been used in conjunction with the two-load or multi-load method to evaluate the source characteristics at a point just downstream of the exhaust manifold for the exhaust system, and upstream of the air filter (dirty side) in the case of the intake system. These source characteristics have been extracted from the pressure time history calculated at that point using the electro-acoustic analogy. Systematic parametric studies have yielded approximate empirical expressions for the source characteristics of an engine in terms of the basic engine parameters like engine RPM, capacity (swept volume or displacement), air-fuel ratio, and the number of cylinders. The effect of other parameters has been found to be relatively insignificant. Unlike exhaust noise, the intake system noise of an automobile cannot be measured because of the proximity of the engine at the point of measurement. Besides, the intake side is associated with turbocharger (booster), intercooler, cooling fan, etc., which will make the measurement of the intake noise erroneous. From the noise radiation point of view, intake noise used to be considered to be a minor source of noise as compared to the exhaust noise. Therefore, very little has been done or reported on prediction of the intake noise as compared to the exhaust noise. But nowadays, with efficient exhaust mufflers, the un-muffled intake noise has become a contributing factor to the passenger compartment noise level as a luxury decisive factor. Therefore, in this investigation both the intake and the exhaust side source characteristics have been found out for the compression ignition as well as the spark ignition engines. Besides, in the case of compression ignition engines, typical turbocharged as well as naturally aspirated engines have been considered. One of the inputs to the time-domain simulation is the intake valve and exhaust valve lift histories as functions of crank angle. It is very cumbersome and time-consuming to measure and feed these data into the program. Sometimes, this data is not available or cannot be determined easily. So, a generalized formula for the valve lift has been developed by observing the valve lift curves of various engines. The maximum exhaust valve lift has been expressed as a function of the swept volume of the cylinder. This formulation is not intended for designing a cam profile; it is for the purpose of determining approximate thermodynamic quantities to help a muffler designer for an initial estimation. It has also been observed during the investigation that from the acoustic point of view, sometimes it is better to open the exhaust valve a little earlier, but very slowly and smoothly, and keep it open for a longer time. Although the exact source characteristics for an automobile engine cannot be determined precisely, yet the values of source characteristics calculated using this methodology have been shown to be reasonably good for approximate prediction of the un-muffled noise as well as insertion loss of a given muffler. The resultant empirical expressions for the source characteristics enable the potential user to make use of the frequency-domain cum-transfer matrix approach throughout; the time consuming time-domain simulation of the engine exhaust source is no longer necessary. Predictions of the un-muffled sound pressure level of automotive engines have been corroborated against measured values as the well as the full scale time-domain predictions making use of a finite-volume software.