Lip-reading is mostly used as a means of communication by people with hearing di�fficulties. Recent work has explored the automation of this process, with the aim of building a speech recognition system entirely driven by lip movements. However, this work has so far produced poor results because of factors such as high variability of speaker features, diffi�culties in mapping from visual features to speech sounds, and high co-articulation of visual features. The motivation for the work in this thesis is inspired by previous work in dysarthric speech recognition [Morales, 2009]. Dysathric speakers have poor control over their articulators, often leading to a reduced phonemic repertoire. The premise of this thesis is that recognition of the visual speech signal is a similar problem to recog- nition of dysarthric speech, in that some information about the speech signal has been lost in both cases, and this brings about a systematic pattern of errors in the decoded output. This work attempts to exploit the systematic nature of these errors by modelling them in the framework of a weighted finite-state transducer cascade. Results indicate that the technique can achieve slightly lower error rates than the conventional approach. In addition, it explores some interesting more general questions for automated lip-reading.

Acentrosomal microtubules are not bound to a microtubule organising centre yet are still able to form ordered arrays. Two clear examples of this behaviour are the acentrosomal apico-basal (side wall) array in epithelial cells and the parallel organisation of plant cortical microtubules. This research investigates their formation through mathematical modelling and Monte Carlo simulations with the software programs developed ourselves. In epithelial cells there is a generally accepted `release and capture' hypothesis for the transfer of centrosomal microtubules onto the side wall array. We use a combination of mathematical and Monte Carlo simulation models to perform the first modelling of this hypothesis. We find that a tubulin concentration dependent dynamic instability is not a good�fit to this hypothesis but that a reduced centrosomal nucleation rate in response to an increased number of side wall microtubules makes the hypothesis work in biologically reasonable conditions. We propose that the loss of nucleation rate is a result of ninein being transferred from the centrosome to the side wall. We show OpenCL to be a useful tool in building a simulation program for parameter searches. We use a Monte Carlo simulation model to investigate how the collision induced catastrophe (CIC) probability affects the formation of the ordered array of cortical plant microtubules. We find that with entrainment an ordered array stops forming once the CIC drops below 0.5. We�find that the severing action of katanin is able to restore order at CIC probabilities below 0.5 but the speed at which crossovers must be severed becomes unfeasibly fast as the CIC decreases. This implies that at very low CICs observed in nature (�0.1), katanin may be necessary but not suffi�cient to create the ordered array. We also provide a customisable and intuitive cortical microtubule simulation software to aid in further research.

There has been considerable research on the problems of speaker and language recognition from samples of speech. A less researched problem is that of accent recognition. Although this is a similar problem to language identification, different accents of a language exhibit more fine-grained differences between classes than languages. This presents a tougher problem for traditional classification techniques. In this thesis, we propose and evaluate a number of techniques for gender and accent classification. These techniques are novel modifications and extensions to state of the art algorithms, and they result in enhanced performance on gender and accent recognition. The first part of the thesis focuses on the problem of gender identification, and presents a technique that gives improved performance in situations where training and test conditions are mismatched. The bulk of this thesis is concerned with the application of the i-Vector technique to accent identification, which is the most successful approach to acoustic classification to have emerged in recent years. We show that it is possible to achieve high accuracy accent identification without reliance on transcriptions and without utilising phoneme recognition algorithms. The thesis describes various stages in the development of i-Vector based accent classification that improve the standard approaches usually applied for speaker or language identification, which are insufficient. We demonstrate that very good accent identification performance is possible with acoustic methods by considering different i-Vector projections, frontend parameters, i-Vector configuration parameters, and an optimised fusion of the resulting i-Vector classifiers we can obtain from the same data. We claim to have achieved the best accent identification performance on the test corpus for acoustic methods, with up to 90% identification rate. This performance is even better than previously reported acoustic-phonotactic based systems on the same corpus, and is very close to performance obtained via transcription based accent identification. Finally, we demonstrate that the utilization of our techniques for speech recognition purposes leads to considerably lower word error rates.

No description available.

The advent of shaders in the latest generations of graphics hardware, which has made consumer level graphics hardware partially programmable, makes now an ideal time to investigate new graphical techniques and algorithms as well as attempting to improve upon existing ones. This work looks at areas of current interest within the graphics community such as Texture Filtering, Bump Mapping and Depth of Field simulation. These are all areas which have enjoyed much interest over the history of computer graphics but which provide a great deal of scope for further investigation in the light of recent hardware advances. A new hardware implementation of a texture filtering technique, aimed at consumer level hardware, is presented. This novel technique utilises Fourier space image filtering to reduce aliasing. Investigation shows that the technique provides reduced levels of aliasing along with comparable levels of detail to currently popular techniques. This adds to the community's knowledge by expanding the range of techniques available, as well as increasing the number of techniques which offer the potential for easy integration with current consumer level graphics hardware along with real-time performance. Bump mapping is a long-standing and well understood technique. Variations and extensions of it have been popular in real-time 3D computer graphics for many years. A new hardware implementation of a technique termed Super Bump Mapping (SBM) is introduced. Expanding on the work of Cant and Langensiepen [1], the SBM technique adopts the novel approach of using normal maps which supply multiple vectors per texel. This allows the retention of much more detail and overcomes some of the aliasing deficiencies of standard bump mapping caused by the standard single vector approach and the non-linearity of the bump mapping process. A novel depth of field algorithm is proposed, which is an extension of the authors previous work [2][3][4]. The technique is aimed at consumer level hardware and attempts to raise the bar for realism by providing support for the 'see-through' effect. This effect is a vital factor in the realistic appearance of simulated depth of field and has been overlooked in real time computer graphics due to the complexities of an accurate calculation. The implementation of this new algorithm on current consumer level hardware is investigated and it is concluded that while current hardware is not yet capable enough, future iterations will provide the necessary functional and performance increases.

High performance architectures are increasingly heterogeneous with shared and distributed memory components, and accelerators like GPUs. Programming such architectures is complicated and performance portability is a major issue as the architectures evolve. This thesis explores the potential for algorithmic skeletons integrating a dynamically parametrised static cost model, to deliver portable performance for mostly regular data parallel programs on heterogeneous archi- tectures. The rst contribution of this thesis is to address the challenges of program- ming heterogeneous architectures by providing two skeleton-based programming libraries: i.e. HWSkel for heterogeneous multicore clusters and GPU-HWSkel that enables GPUs to be exploited as general purpose multi-processor devices. Both libraries provide heterogeneous data parallel algorithmic skeletons including hMap, hMapAll, hReduce, hMapReduce, and hMapReduceAll. The second contribution is the development of cost models for workload dis- tribution. First, we construct an architectural cost model (CM1) to optimise overall processing time for HWSkel heterogeneous skeletons on a heterogeneous system composed of networks of arbitrary numbers of nodes, each with an ar- bitrary number of cores sharing arbitrary amounts of memory. The cost model characterises the components of the architecture by the number of cores, clock speed, and crucially the size of the L2 cache. Second, we extend the HWSkel cost model (CM1) to account for GPU performance. The extended cost model (CM2) is used in the GPU-HWSkel library to automatically nd a good distribution for both a single heterogeneous multicore/GPU node, and clusters of heteroge- neous multicore/GPU nodes. Experiments are carried out on three heterogeneous multicore clusters, four heterogeneous multicore/GPU clusters, and three single heterogeneous multicore/GPU nodes. The results of experimental evaluations for four data parallel benchmarks, i.e. sumEuler, Image matching, Fibonacci, and Matrix Multiplication, show that our combined heterogeneous skeletons and cost models can make good use of resources in heterogeneous systems. Moreover using cores together with a GPU in the same host can deliver good performance either on a single node or on multiple node architectures.

The continuous growth of complex systems makes the development of correct software increasingly challenging. In order to address this challenge, formal methods o er rigorous mathematical techniques to model and verify the correctness of systems. Refinement is one of these techniques. By allowing a developer to incrementally introduce design details, refinement provides a powerful mechanism for mastering the complexities that arise when formally modelling systems. Here the focus is on a posit-and-prove style of refinement, where a design is developed as a series of abstract models introduced via refinement steps. Each refinement step generates proof obligations which must be discharged in order to verify its correctness – typically requiring a user to understand the relationship between modelling and reasoning. This thesis focuses on techniques to aid refinement-based formal modelling, specifically, when a user requires guidance in order to overcome a failed refinement step. An integrated approach has been followed: combining the complementary strengths of bottomup theory formation, in which theories about domains are built based on basic background information; and top-down planning, in which meta-level reasoning is used to guide the search for correct models. On the theory formation perspective, we developed a technique for the automatic discovery of invariants. Refinement requires the definition of properties, called invariants, which relate to the design. Formulating correct and meaningful invariants can be tedious and a challenging task. A heuristic approach to the automatic discovery of invariants has been developed building upon simulation, proof-failure analysis and automated theory formation. This approach exploits the close interplay between modelling and reasoning in order to provide systematic guidance in tailoring the search for invariants for a given model. On the planning perspective, we propose a new technique called refinement plans. Refinement plans provide a basis for automatically generating modelling guidance when a step fails but is close to a known pattern of refinement. This technique combines both modelling and reasoning knowledge, and, contrary to traditional pattern techniques, allow the analysis of failure and partial matching. Moreover, when the guidance is only partially instantiated, and it is suitable, refinement plans provide specialised knowledge to further tailor the theory formation process in an attempt to fully instantiate the guidance. We also report on a series of experiments undertaken in order to evaluate the approaches and on the implementation of both techniques into prototype tools. We believe the techniques presented here allow the developer to focus on design decisions rather than on analysing low-level proof failures.

Blind source separation (BSS) techniques have the aim of separating original source signals from their mixtures without having or with a little knowledge about the source signals or the mixing process. Tensor based source separation techniques have become increasingly popular for various applications since they exploit different inherent diversities of the sources. Therefore, they can improve the estimation of desired sources and the identification of the mixing system. The proposed techniques in this thesis are the extensions of conventional tensor factorisation techniques. Our proposed methods can be categorised in two groups; single and multichannel source separation techniques. For single channel source separation a tensor based singular spectrum analysis (SSA) is proposed followed by a way to select the desired subspaces automatically. The proposed method is compared with conventional methods using both synthetic and real electroencephalography (EEG) sleep data to track different sleep stages. Another proposed method is symmetric tensor decomposition. The method has been applied to detect the beta rebound, as an indicator of movement related brain responses, in brain computer interfacing (BCI). In addition to single channel source separation, several multichannel BSS techniques have been proposed. The first method is a constrained BSS technique which uses the spatial information of data to improve the performance. This approach is particularly useful in separation of weak intermittent signal components. The results show that the proposed method performs better than the existing methods in terms of accuracy and quality. Complex tensor factorisation of correlated brain sources is attempted as well. The electrical brain responses due to motory, sensory, or cognitive stimuli, i.e. event related potentials (ERPs), particularly P300, have been used for cognitive information processing. P300 has two subcomponents, P3a and P3b, which are correlated, and therefore, the traditional BSS approaches fail in their separation. A complex-valued tensor factorisation of EEG signals is introduced with the aim of separating P300 subcomponents. The proposed method uses complex-valued statistics to exploit the data correlation. In this way, the variations of P3a and p3b can be tracked for the assessment of the brain state. The results of this work are compared with those of spatial principle component analysis (SPCA) method. Communication signals such as quadrature-phase shift keying (QPSK) often pose as complex waveforms and suffer from multipath and clutter problems. In this thesis therefore, a new convolutive complex tensor factorisation system is proposed to recover such signals in the receiver and estimate the communication channels. The proposed method is evaluated using simulated data with multiple multi-paths and various non-circularity and noise levels. Simulation results confirm the superiority of the proposed method over the existing popular techniques.

This thesis studies two major extensions of nominal terms. In particular, we study an extension with -abstraction over nominal unknowns and atoms, and an extension with an arguably better theory of freshness and -equivalence. Nominal terms possess two levels of variable: atoms a represent variable symbols, and unknowns X are `real' variables. As a syntax, they are designed to facilitate metaprogramming; unknowns are used to program on syntax with variable symbols. Originally, the role of nominal terms was interpreted narrowly. That is, they were seen solely as a syntax for representing partially-speci ed abstract syntax with binding. The main motivation of this thesis is to extend nominal terms so that they can be used for metaprogramming on proofs, programs, etc. and not just for metaprogramming on abstract syntax with binding. We therefore extend nominal terms in two signi cant ways: adding -abstraction over nominal unknowns and atoms| facilitating functional programing|and improving the theory of -equivalence that nominal terms possesses. Neither of the two extensions considered are trivial. The capturing substitution action of nominal unknowns implies that our notions of scope, intuited from working with syntax possessing a non-capturing substitution, such as the -calculus, is no longer applicable. As a result, notions of -abstraction and -equivalence must be carefully reconsidered. In particular, the rst research contribution of this thesis is the two-level - calculus, intuitively an intertwined pair of -calculi. As the name suggests, the two-level -calculus has two level of variable, modelled by nominal atoms and unknowns, respectively. Both levels of variable can be -abstracted, and requisite notions of -reduction are provided. The result is an expressive context-calculus. The traditional problems of handling -equivalence and the failure of commutation between instantiation and -reduction in context-calculi are handled through the use of two distinct levels of variable, swappings, and freshness side-conditions on unknowns, i.e. `nominal technology'. The second research contribution of this thesis is permissive nominal terms, an alternative form of nominal term. They retain the `nominal' rst-order avour of nominal terms (in fact, their grammars are almost identical) but forego the use of explicit freshness contexts. Instead, permissive nominal terms label unknowns with a permission sort, where permission sorts are in nite and coin nite sets of atoms. This in nite-coin nite nature means that permissive nominal terms recover two properties|we call them the `always-fresh' and `always-rename' properties that nominal terms lack. We argue that these two properties bring the theory of -equivalence on permissive nominal terms closer to `informal practice'. The reader may consider -abstraction and -equivalence so familiar as to be `solved problems'. The work embodied in this thesis stands testament to the fact that this isn't the case. Considering -abstraction and -equivalence in the context of two levels of variable poses some new and interesting problems and throws light on some deep questions related to scope and binding.

The proliferation of freely-accessible data-intensive websites, the growing availability of pervasive and mobile applications, as well as initiatives for open-accessible linked data in the Web, provided the users with potential sources of valuable information. These data also represent new business opportunities for industries, while their processing and management is a rich research field for academics. In addition, pervasiveness and mobility make information available everywhere and at any time; therefore, user-centred, dynamic and on-the-fly integration of heterogeneous becomes extremely useful in practical situations such as mobile data management, in particular with devices as tablets, pads and smartphones. We present a methodology and a set of technologies for ontology-driven, context-aware data-integration systems capable of seamlessly handling heterogeneous and dynamic data sources in a dynamic environment where the possibly mobile, transient and evolving data sources are not known in advance. In such Nomadic Data-Integration Systems (N-DIS), query answering is driven by an ontology and the context is used to formally represent the situations in which the users or the applications need to operate. In our work, we re-think query-answering techniques coming from description logics and database theory, and we apply them to an evolving scenario, composed by heterogeneous data sources with the additional constraint of the context-based personalisation of queries and, as a consequence, of their answers. The concept of N-DIS and the corresponding design methodology has been applied to various application fields such as decision support for medical-emergency situations, Semantic-Web meta-data management, Semantic Search and pervasive advertisement, with promising results.