Improving short utterance i-vector speaker verification using utterance variance modelling and compensation techniques

This paper proposes techniques to improve the performance of i-vector based speaker verification systems when only short utterances are available. Short-length utterance i-vectors vary with speaker, session variations, and the phonetic content of the utterance. Well established methods such as linear discriminant analysis (LDA), source-normalized LDA (SN-LDA) and within-class covariance normalisation (WCCN) exist for compensating the session variation but we have identified the variability introduced by phonetic content due to utterance variation as an additional source of degradation when short-duration utterances are used. To compensate for utterance variations in short i-vector speaker verification systems using cosine similarity scoring (CSS), we have introduced a short utterance variance normalization (SUVN) technique and a short utterance variance (SUV) modelling approach at the i-vector feature level. A combination of SUVN with LDA and SN-LDA is proposed to compensate the session and utterance variations and is shown to provide improvement in performance over the traditional approach of using LDA and/or SN-LDA followed by WCCN. An alternative approach is also introduced using probabilistic linear discriminant analysis (PLDA) approach to directly model the SUV. The combination of SUVN, LDA and SN-LDA followed by SUV PLDA modelling provides an improvement over the baseline PLDA approach. We also show that for this combination of techniques, the utterance variation information needs to be artificially added to full-length i-vectors for PLDA modelling.

Improving PLDA speaker verification with limited development data

This paper analyses the probabilistic linear discriminant analysis (PLDA) speaker verification approach with limited development data. This paper investigates the use of the median as the central tendency of a speaker’s i-vector representation, and the effectiveness of weighted discriminative techniques on the performance of state-of-the-art length-normalised Gaussian PLDA (GPLDA) speaker verification systems. The analysis within shows that the median (using a median fisher discriminator (MFD)) provides a better representation of a speaker when the number of representative i-vectors available during development is reduced, and that further, usage of the pair-wise weighting approach in weighted LDA and weighted MFD provides further improvement in limited development conditions. Best performance is obtained using a weighted MFD approach, which shows over 10% improvement in EER over the baseline GPLDA system on mismatched and interview-interview conditions.

Experimental verification of the modified spring-mass theory of fiber Bragg grating accelerometers using transverse forces

A fiber Bragg grating (FBG) accelerometer using transverse forces is more sensitive than one using axial forces with the same mass of the inertial object, because a barely stretched FBG fixed at its two ends is much more sensitive to transverse forces than axial ones. The spring-mass theory, with the assumption that the axial force changes little during the vibration, cannot accurately predict its sensitivity and resonant frequency in the gravitational direction because the assumption does not hold due to the fact that the FBG is barely prestretched. It was modified but still required experimental verification due to the limitations in the original experiments, such as the (1) friction between the inertial object and shell; (2) errors involved in estimating the time-domain records; (3) limited data; and (4) large interval ∼5 Hz between the tested frequencies in the frequency-response experiments. The experiments presented here have verified the modified theory by overcoming those limitations. On the frequency responses, it is observed that the optimal condition for simultaneously achieving high sensitivity and resonant frequency is at the infinitesimal prestretch. On the sensitivity at the same frequency, the experimental sensitivities of the FBG accelerometer with a 5.71 gram inertial object at 6 Hz (1.29, 1.19, 0.88, 0.64, and 0.31 nm/g at the 0.03, 0.69, 1.41, 1.93, and 3.16 nm prestretches, respectively) agree with the static sensitivities predicted (1.25, 1.14, 0.83, 0.61, and 0.29 nm/g, correspondingly). On the resonant frequency, (1) its assumption that the resonant frequencies in the forced and free vibrations are similar is experimentally verified; (2) its dependence on the distance between the FBG’s fixed ends is examined, showing it to be independent; (3) the predictions of the spring-mass theory and modified theory are compared with the experimental results, showing that the modified theory predicts more accurately. The modified theory can be used more confidently in guiding its design by predicting its static sensitivity and resonant frequency, and may have applications in other fields for the scenario where the spring-mass theory fails.

Connectivity of workflow nets : the foundations of stepwise verification

Behavioral models capture operational principles of real-world or designed systems. Formally, each behavioral model defines the state space of a system, i.e., its states and the principles of state transitions. Such a model is the basis for analysis of the system’s properties. In practice, state spaces of systems are immense, which results in huge computational complexity for their analysis. Behavioral models are typically described as executable graphs, whose execution semantics encodes a state space. The structure theory of behavioral models studies the relations between the structure of a model and the properties of its state space. In this article, we use the connectivity property of graphs to achieve an efficient and extensive discovery of the compositional structure of behavioral models; behavioral models get stepwise decomposed into components with clear structural characteristics and inter-component relations. At each decomposition step, the discovered compositional structure of a model is used for reasoning on properties of the whole state space of the system. The approach is exemplified by means of a concrete behavioral model and verification criterion. That is, we analyze workflow nets, a well-established tool for modeling behavior of distributed systems, with respect to the soundness property, a basic correctness property of workflow nets. Stepwise verification allows the detection of violations of the soundness property by inspecting small portions of a model, thereby considerably reducing the amount of work to be done to perform soundness checks. Besides formal results, we also report on findings from applying our approach to an industry model collection.

The biconnected verification of workflow nets

Formal representations of business processes are used for analysis of the process behavior. Workflow nets are a widely used formalism for describing the behavior of business processes. Structure theory of processes investigates the relation between the structure of a model and its behavior. In this paper, we propose to employ the connectivity property of workflow nets as an angle to their structural analysis. In particular, we show how soundness verification can be organized using biconnected components of a workflow net. This allows for efficient identification and localization of flaws in the behavior of workflow nets and for supporting process analysts with diagnostic information

Three-dimensional dose verification for clinical treatments of small intracranial tumours

Cancers of the brain and central nervous system account for 1.6% of new cancers and 1.8% of cancer deaths globally. The highest rates of all developed nations are observed in Australia and New Zealand. There are known complexities associated with dose measurement of very small radiation fields. Here, 3D dosimetric verification of treatments for small intracranial tumours using gel dosimetry was investigated.

Generation of uniform plasmas by crossed internal oscillating current sheets : key concepts and experimental verification

The results of comprehensive experimental studies of the operation, stability, and plasma parameters of the low-frequency (0.46 MHz) inductively coupled plasmas sustained by the internal oscillating rf current are reported. The rf plasma is generated by using a custom-designed configuration of the internal rf coil that comprises two perpendicular sets of eight currents in each direction. Various diagnostic tools, such as magnetic probes, optical emission spectroscopy, and an rf-compensated Langmuir probe were used to investigate the electromagnetic, optical, and global properties of the argon plasma in wide ranges of the applied rf power and gas feedstock pressure. It is found that the uniformity of the electromagnetic field inside the plasma reactor is improved as compared to the conventional sources of inductively coupled plasmas with the external flat coil configuration. A reasonable agreement between the experimental data and computed electromagnetic field topography inside the chamber is reported. The Langmuir probe measurements reveal that the spatial profiles of the electron density, the effective electron temperature, plasma potential, and electron energy distribution/probability functions feature a high degree of the radial and axial uniformity and a weak azimuthal dependence, which is consistent with the earlier theoretical predictions. As the input rf power increases, the azimuthal dependence of the global plasma parameters vanishes. The obtained results demonstrate that by introducing the internal oscillated rf currents one can noticeably improve the uniformity of electromagnetic field topography, rf power deposition, and the plasma density in the reactor.

Short utterance PLDA speaker verification using SN-WLDA and variance modelling techniques

This paper proposes a combination of source-normalized weighted linear discriminant analysis (SN-WLDA) and short utterance variance (SUV) PLDA modelling to improve the short utterance PLDA speaker verification. As short-length utterance i-vectors vary with the speaker, session variations and phonetic content of the utterance (utterance variation), a combined approach of SN-WLDA projection and SUV PLDA modelling is used to compensate the session and utterance variations. Experimental studies have found that a combination of SN-WLDA and SUV PLDA modelling approach shows an improvement over baseline system (WCCN[LDA]-projected Gaussian PLDA (GPLDA)) as this approach effectively compensates the session and utterance variations.

A runtime integrity monitoring framework for real-time relative positioning systems based on GPS and DSRC

This paper provides a three-layered framework to monitor the positioning performance requirements of Real-time Relative Positioning (RRP) systems of the Cooperative Intelligent Transport Systems (C-ITS) that support Cooperative Collision Warning (CCW) applications. These applications exploit state data of surrounding vehicles obtained solely from the Global Positioning System (GPS) and Dedicated Short-Range Communications (DSRC) units without using other sensors. To this end, the paper argues the need for the GPS/DSRC-based RRP systems to have an autonomous monitoring mechanism, since the operation of CCW applications is meant to augment safety on roads. The advantages of autonomous integrity monitoring are essential and integral to any safety-of-life system. The autonomous integrity monitoring framework proposed necessitates the RRP systems to detect/predict the unavailability of their sub-systems and of the integrity monitoring module itself, and, if available, to account for effects of data link delays and breakages of DSRC links, as well as of faulty measurement sources of GPS and/or integrated augmentation positioning systems, before the information used for safety warnings/alarms becomes unavailable, unreliable, inaccurate or misleading. Hence, a monitoring framework using a tight integration and correlation approach is proposed for instantaneous reliability assessment of the RRP systems. Ultimately, using the proposed framework, the RRP systems will provide timely alerts to users when the RRP solutions cannot be trusted or used for the intended operation.

Sensitivity of an Elekta iView GT a-Si EPID model to delivery errors for pre-treatment verification of IMRT fields

A Monte Carlo model of an Elekta iViewGT amorphous silicon electronic portal imaging device (a-Si EPID) has been validated for pre-treatment verification of clinical IMRT treatment plans. The simulations involved the use of the BEAMnrc and DOSXYZnrc Monte Carlo codes to predict the response of the iViewGT a-Si EPID model. The predicted EPID images were compared to the measured images obtained from the experiment. The measured EPID images were obtained by delivering a photon beam from an Elekta Synergy linac to the Elekta iViewGT a-Si EPID. The a-Si EPID was used with no additional build-up material. Frame averaged EPID images were acquired and processed using in-house software. The agreement between the predicted and measured images was analyzed using the gamma analysis technique with acceptance criteria of 3% / 3 mm. The results show that the predicted EPID images for four clinical IMRT treatment plans have a good agreement with the measured EPID signal. Three prostate IMRT plans were found to have an average gamma pass rate of more than 95.0 % and a spinal IMRT plan has the average gamma pass rate of 94.3 %. During the period of performing this work a routine MLC calibration was performed and one of the IMRT treatments re-measured with the EPID. A change in the gamma pass rate for one field was observed. This was the motivation for a series of experiments to investigate the sensitivity of the method by introducing delivery errors, MLC position and dosimetric overshoot, into the simulated EPID images. The method was found to be sensitive to 1 mm leaf position errors and 10% overshoot errors.

Speaker verification using I-vector features

This PhD research has provided novel solutions to three major challenges which have prevented the wide spread deployment of speaker recognition technology: (1) combating enrolment/ verification mismatch, (2) reducing the large amount of development and training data that is required and (3) reducing the duration of speech required to verify a speaker. A range of applications of speaker recognition technology from forensics in criminal investigations to secure access in banking will benefit from the research outcomes.

Veracity of the archive : a research approach to the collection and verification of urban morphological records using qualitative data analysis software

Critical to the research of urban morphologists is the availability of historical records that document the urban transformation of the study area. However, thus far little work has been done towards an empirical approach to the validation of archival data in this field. Outlined in this paper, therefore, is a new methodology for validating the accuracy of archival records and mapping data, accrued through the process of urban morphological research, so as to establish a reliable platform from which analysis can proceed. The paper particularly addresses the problems of inaccuracies in existing curated historical information, as well as errors in archival research by student assistants, which together give rise to unacceptable levels of uncertainty in the documentation. The paper discusses the problems relating to the reliability of historical information, demonstrates the importance of data verification in urban morphological research, and proposes a rigorous method for objective testing of collected archival data through the use of qualitative data analysis software.

Dynamic agent composition for large-scale agent-based models

PURPOSE: This paper describes dynamic agent composition, used to support the development of flexible and extensible large-scale agent-based models (ABMs). This approach was motivated by a need to extend and modify, with ease, an ABM with an underlying networked structure as more information becomes available. Flexibility was also sought after so that simulations are set up with ease, without the need to program. METHODS: The dynamic agent composition approach consists in having agents, whose implementation has been broken into atomic units, come together at runtime to form the complex system representation on which simulations are run. These components capture information at a fine level of detail and provide a vast range of combinations and options for a modeller to create ABMs. RESULTS: A description of the dynamic agent composition is given in this paper, as well as details about its implementation within MODAM (MODular Agent-based Model), a software framework which is applied to the planning of the electricity distribution network. Illustrations of the implementation of the dynamic agent composition are consequently given for that domain throughout the paper. It is however expected that this approach will be beneficial to other problem domains, especially those with a networked structure, such as water or gas networks. CONCLUSIONS: Dynamic agent composition has many advantages over the way agent-based models are traditionally built for the users, the developers, as well as for agent-based modelling as a scientific approach. Developers can extend the model without the need to access or modify previously written code; they can develop groups of entities independently and add them to those already defined to extend the model. Users can mix-and-match already implemented components to form large-scales ABMs, allowing them to quickly setup simulations and easily compare scenarios without the need to program. The dynamic agent composition provides a natural simulation space over which ABMs of networked structures are represented, facilitating their implementation; and verification and validation of models is facilitated by quickly setting up alternative simulations.

Improving out-domain PLDA speaker verification using unsupervised inter-dataset variability compensation approach

Experimental studies have found that when the state-of-the-art probabilistic linear discriminant analysis (PLDA) speaker verification systems are trained using out-domain data, it significantly affects speaker verification performance due to the mismatch between development data and evaluation data. To overcome this problem we propose a novel unsupervised inter dataset variability (IDV) compensation approach to compensate the dataset mismatch. IDV-compensated PLDA system achieves over 10% relative improvement in EER values over out-domain PLDA system by effectively compensating the mismatch between in-domain and out-domain data.