For the last several years, speaker diarization has been attracting substantial research attention as one of the spoken language technologies applied for the improvement, or enrichment, of recording transcriptions. Recordings of meetings, compared to other domains, exhibit an increased complexity due to the spontaneity of speech, reverberation effects, and also due to the presence of overlapping speech. Overlapping speech refers to situations when two or more speakers are speaking simultaneously. In meeting data, a substantial portion of errors of the conventional speaker diarization systems can be ascribed to speaker overlaps, since usually only one speaker label is assigned per segment. Furthermore, simultaneous speech included in training data can eventually lead to corrupt single-speaker models and thus to a worse segmentation. This thesis concerns the detection of overlapping speech segments and its further application for the improvement of speaker diarization performance. We propose the use of three spatial cross-correlationbased parameters for overlap detection on distant microphone channel data. Spatial features from different microphone pairs are fused by means of principal component analysis, linear discriminant analysis, or by a multi-layer perceptron. In addition, we also investigate the possibility of employing longterm prosodic information. The most suitable subset from a set of candidate prosodic features is determined in two steps. Firstly, a ranking according to mRMR criterion is obtained, and then, a standard hill-climbing wrapper approach is applied in order to determine the optimal number of features. The novel spatial as well as prosodic parameters are used in combination with spectral-based features suggested previously in the literature. In experiments conducted on AMI meeting data, we show that the newly proposed features do contribute to the detection of overlapping speech, especially on data originating from a single recording site. In speaker diarization, for segments including detected speaker overlap, a second speaker label is picked, and such segments are also discarded from the model training. The proposed overlap labeling technique is integrated in Viterbi decoding, a part of the diarization algorithm. During the system development it was discovered that it is favorable to do an independent optimization of overlap exclusion and labeling with respect to the overlap detection system. We report improvements over the baseline diarization system on both single- and multi-site AMI data. Preliminary experiments with NIST RT data show DER improvement on the RT ¿09 meeting recordings as well. The addition of beamforming and TDOA feature stream into the baseline diarization system, which was aimed at improving the clustering process, results in a bit higher effectiveness of the overlap labeling algorithm. A more detailed analysis on the overlap exclusion behavior reveals big improvement contrasts between individual meeting recordings as well as between various settings of the overlap detection operation point. However, a high performance variability across different recordings is also typical of the baseline diarization system, without any overlap handling.

In the current study, the four most common Particle Image Velocimetry (PIV) cross-correlation algorithms were evaluated by measuring the displacement of particles in computer generated images. The synthetic images were employed to compare the methods since the particle diameter, density, and intensity could be controlled, removing some of the uncertainty found in images collected during experiments, e. g. parallax, 3-D motion, etc. The most important parameter that was controlled in the synthetic images was the particle motion. Six different displacement functions were applied to move the particles between images: uniform translation, step, sawtooth, sinusoid, line source and line vortex. The four algorithms, which all use the fast Fourier transform (FFT) to perform the cross-correlation, were evaluated with four criteria; (1) spatial resolution, (2) dynamic range, (3) accuracy and (4) robustness. The uniform translation images determined the least error possible with each method, of which the deformed FFT proved to be the most accurate. The super resolution FFT and deformed FFT methods could not properly measure the infinite displacement gradient in the step images due to the interpolation of the displacement vector field used by each method around the step. However, the predictor corrector FFT scheme, which does not require interpolation when determining the interrogation area offset, successfully measured the infinite displacement gradient in the step images. The smaller interrogation areas used by the super resolution FFT scheme proved to be the best method to capture the high frequency finite displacement gradients in the sawtooth and sinusoid images. Also shown in the sawtooth and sinusoid images is the positional bias error introduced by assuming the measured particle displacement occurs at the centre of the interrogation area. The deformed FFT method produced the most accurate results for the source and vortex images, which both contained displacement gradients in multiple directions. Experimentally obtained images were also evaluated to verify the results derived using the synthetic images. The flow in a multiple grooved channel, using both water and air as the fluid medium in separate experiments, was measured and compared to DNS simulations reported by Yang. The mean velocity, average vorticity and turbulent fluctuations determined from both experiments using the deformed FFT method compared very well to the DNS calculations.

The Sun’s corona expands outward, populating the solar system with plasma. This plasma is known as the solar wind. The solar wind carries with it the Sun’s magnetic field, which is also known as the interplanetary magnetic field (IMF). The resulting configuration of the IMF creates a current sheet at solar equatorial latitudes, which the Earth crosses as it orbits the Sun. When the Earth is on one side of the current sheet it is in a sector where the IMF is directed largely away from or toward the Sun. On the other side of the current sheet the IMF is in opposite direction. The crossing of the current sheet is known as a sector boundary crossing (SBC). The solar wind and IMF properties change significantly near the current sheet, and this affects the Earth’s ionosphere. The Super Dual Auroral Radar Network (SuperDARN) high frequency (HF) radar data rates from 2001-2011 were examined using several techniques: a superposed epoch analysis, a fast fourier transform (FFT) analysis, and a cross–correlation analysis. Data from multiple instruments were analyzed in this study. These include the solar wind and IMF data from spacecraft, observations of charged particles precipitating into the Earth’s ionosphere, echoes from ground–based SuperDARN radars, and data from gound–based neutron monitors that detect galactic cosmic rays. Solar wind and IMF properties change significantly across a sector boundary. An increase in the IMF magnitude of about 30% occurs on the day of the SBC, and the IMF returns to pre–crossing values over the next two days. There is a decrease in the solar wind speed of about 15% the day before and the day of the SBC, and the solar wind density doubles at the time of the SBC. The polarity of the SBC does not appear to affect the solar wind and IMF. A peak in the data rate of SuperDARN echoes from both the ionosphere and ground occurs within one day of the SBC, though the variability of these data is quite large. The hemispherical power, which is an estimation of the electron energy flux precipitating into the ionosphere derived from satellite observations, increases following a SBC. Satellite particle data also revealed that the equatorward auroral oval boundary moves equatorward following a SBC. The cosmic ray counts at the Earth’s surface appear to be unaffected by the SBC. The solar wind and ionosphere data sets exhibited strong periodicities, and these were harmonics of the synodic rotational period of the Sun (approximately 27 days). Common periodicities observed were 27 days, 13.5 days, 9 days, 6.75 days and 5.4 days. There was a dominant 9–day periodicity observed in the solar wind and ionospheric data from 2005–2008, but was not observed in the solar 10.7 cm wavelength electromagnetic flux. The 9-day periodicity in the solar wind during this period has been linked to three persistent features on the Sun that produced corotating high–speed streams, or areas of fast solar wind. The parameters whose change did not depend on the polarity of the SBC had periodicities that were half that of the SBCs. From the cross–correlation analysis some relationships between the data sets became evident. For periods of high solar wind speed there were low SuperDARN data rates, and vice versa. The solar wind speed and hemispherical power were found to be well correlated, while the hemispherical power and the SuperDARN scatter occurrence were found to be anticorrelated. The solar wind changes appear to be affecting the state of the ionosphere, likely through particle precipitation. The SuperDARN scatter occurrence has been shown in past studies to be most greatly affected by changes in the electron density profile of the ionosphere, which can be influenced by changes in particle precipitation. These results demonstrate a link between the solar wind and the state of the ionosphere.

In the current study, the four most common Particle Image Velocimetry (PIV) cross-correlation algorithms were evaluated by measuring the displacement of particles in computer generated images. The synthetic images were employed to compare the methods since the particle diameter, density, and intensity could be controlled, removing some of the uncertainty found in images collected during experiments, e. g. parallax, 3-D motion, etc. The most important parameter that was controlled in the synthetic images was the particle motion. Six different displacement functions were applied to move the particles between images: uniform translation, step, sawtooth, sinusoid, line source and line vortex. The four algorithms, which all use the fast Fourier transform (FFT) to perform the cross-correlation, were evaluated with four criteria; (1) spatial resolution, (2) dynamic range, (3) accuracy and (4) robustness. The uniform translation images determined the least error possible with each method, of which the deformed FFT proved to be the most accurate. The super resolution FFT and deformed FFT methods could not properly measure the infinite displacement gradient in the step images due to the interpolation of the displacement vector field used by each method around the step. However, the predictor corrector FFT scheme, which does not require interpolation when determining the interrogation area offset, successfully measured the infinite displacement gradient in the step images. The smaller interrogation areas used by the super resolution FFT scheme proved to be the best method to capture the high frequency finite displacement gradients in the sawtooth and sinusoid images. Also shown in the sawtooth and sinusoid images is the positional bias error introduced by assuming the measured particle displacement occurs at the centre of the interrogation area. The deformed FFT method produced the most accurate results for the source and vortex images, which both contained displacement gradients in multiple directions. Experimentally obtained images were also evaluated to verify the results derived using the synthetic images. The flow in a multiple grooved channel, using both water and air as the fluid medium in separate experiments, was measured and compared to DNS simulations reported by Yang. The mean velocity, average vorticity and turbulent fluctuations determined from both experiments using the deformed FFT method compared very well to the DNS calculations.

Blind source separation methods such as independent component analysis (ICA) and second order blind identification (SOBI) have shown considerable potential in the area of ambient vibration system identification. The objective of these methods is to separate the modal responses, or sources, from the measured output responses, without the knowledge of excitation. Several frequency domain and time domain methods have been proposed and successfully implemented in the literature. Whereas frequency-domain methods pose several challenges typical of dealing with signals in the frequency-domain, popular time-domain methods such as NExT/ERA and SSI pose limitations in dealing with noise, low sensor density, modes having low energy content, or in dealing with systems having closely-spaced modes, such as those found in structures with passive energy dissipation devices, for example, tuned mass dampers.Motivated by these challenges, the current research focuses on developing methods to address the problem of separability of sources with low energy content, closely-spaced modes, and under-determined blind identification, that is, when the number of response measurements is less than the number of sources. These methods, requiring the time and frequency diversities of the measured outputs, are referred to as hybrid time and time-frequency source separation methods. The hybrid methods are classified into two categories. In the first one, the basic principles of modified SOBI are extended using the stationary wavelet transform (SWT) in order to improve the separability of sources, thereby improving the quality of identification. In the second category, empirical mode decomposition is employed to extract the intrinsic mode functions from measurements, followed by an estimation of the mode shape matrix using iterative and/or non iterative procedures within the framework of modified-SOBI. Both experimental and large-scale structural simulation results are included to demonstrate the applicability of these hybrid approaches to structural system identification problems.

Purchasing Power Parity (PPP) is an important theory of exchange rate determination. The documents probing into the PPP theory are voluminous nowadays; however, there hasn¡¦t been an agreed conclusion yet. In this paper, we apply the Panel Lagrange Multiplier unit root test, a newly developed panel unit root test that allows for heterogeneous breaks, under both the null and the alternative, in both the level and trend of the series under investigation, addressed by Im, Lee and Tieslau (2010). The validity of PPP theory can be examined by testing the stationary of real exchange rates. We use the data chosen from the countries of Asia, including Taiwan, Japan, Korea, Tailand, Indonesia, Hong Kong and Singapore to proceed the positive analysis. The result shows that no matter we use CPI or WPI as the price index, both considering more about the structure breaks and using the panel unit root test strongly support the PPP theory. And it also shows that when using the WPI as the price index, there would be much more countries support the PPP theory.

A program based on particle tracking velocimetry (PTV) was developed in this work. The program was successfully validated by means of artificial images where parameters such as radius, concentration, and noise were varied in order to test their influence on the results. This program uses the mask cross correlation technique for particle centroid location. The sub-pixel accuracy is achieved using two different methods, the three point Gaussian interpolation method and the center of gravity method. The second method is only used if the first method fails. The object matching algorithm between frames uses cross correlation with a non binarized image. A performance comparison between different particle image velocimetry (PIV) and PTV algorithms was done using the international standard PIV challenge artificial images. The best performance was obtained by the program developed in this work. It showed the best accuracy, and the best spatial resolution by finding the larger number of correct vectors of all algorithm tested. A procedure is proposed to obtain error estimates for real images based on errors calculated with experimental ones. Using this procedure a real PIV image with 20% noise has an estimated average error of 0.1 pixel. Results of the analysis of 200 experimental images are shown for the two best PTV algorithms.

During the past few years there has been a significant interest in spectrum splitting systems to increase the overall efficiency of photovoltaic solar energy systems. However, methods for comparing the performance of spectrum splitting systems and the effects of optical spectral filter design on system performance are not well developed. This dissertation first establishes a method to analyze and compare spectrum splitting systems with different filters, PV cells types and geometries. The method examines the system conversion efficiency in detail and the role of optical spectral filters. A new metric termed the Improvement over Best Bandgap is defined which expresses the efficiency gain of the spectrum splitting system with respect to a similar system that contains the highest constituent single bandgap photovoltaic cell. Also, this work expands the analysis on dispersive spectrum splitting systems. The dispersive effects of transmission type filters are evaluated using a cross-correlation analysis. Lastly, this work presents a grating-over-lens design for dispersive spectrum splitting. In this geometry, a transmission grating is placed at the entrance of a lens. Part of the incident solar spectrum is diffracted off-axis from normal incidence to the lens. The diffracted spectral range comes to a focus at an off-axis point and the undiffracted spectrum comes to a focus at the paraxial focus of the lens. Since the diffracted wave is planar and off-axis, the off-axis focal points suffer from aberrations that increase system loss. In this work, a novel aberration compensation technique is presented using non-planar transmission gratings recorded using a conjugate object beam to modify the off-axis wavefront. Diverging sources are used as conjugate object and reference beams. The spherical waves are incident at the lens and the grating is recorded at the entrance aperture of the solar concentrator. The on-axis source is adjusted to produce an on-axis planar wavefront at the hologram plane. The off-axis source is approximated to a diffraction limited spot producing a non-planar off-axis wavefront on the hologram plane. Illumination with a planar AM1.5 spectrum reproduces an off-axis diffraction-limited spot on the focal plane. Models and experimental data are presented to quantify the reduction in losses achieved with aberration correction.

No description available.

An approximate of the Green's function can be obtained by taking the cross-correlation of ambient noise that has been simultaneously recorded on separate sensors. This method is applied for two experiments, which illustrate the advantages and challenges of this technique. The first experiment is in the ultrasonic regime [5-30] MHz and uses capacitive micromachined ultrasonic transducer arrays to image the near field and compares the passive imaging to the conventional pulse-echo imaging. Both the array and target are immersed in a fluid with the sensors recording the fluid's random thermal-mechanical motion as the ambient noise. The second experiment is a passive ocean monitoring experiment, which uses spatiotemporal filtering to rapidly extract coherent arrivals between two vertical line arrays. In this case the ambient noise in the frequency band [250 1500] Hz is dominated by non-stationary shipping noise. For imaging purposes, the cross-correlation needs to extract the Green's function so that the imaging can be done correctly. While for monitoring purposes, the important feature is the change in arrivals, which corresponds to the environment changing. Results of both experiments are presented along with the advantages of this passive method over the more accepted active methods.