Thesis (M. Phil.)--University of Hong Kong, 1989.

The dissertation presents the design and development of a systematic signal analysis and representation framework beyond short-time Fourier power spectrum for sounds, in particular environmental sounds. This framework is consistent with the underlying assumptions of the analysis method and its elements are correlated with human perception. The sound signal has to conform to certain conditions for its power spectrum to have a physical and perceptual meaning. We contend that very few environmental sounds readily meet these criteria and argue that the quantities that are traditionally used to describe sounds need to be repurposed and, if necessary, redefined to represent sounds by non-Fourier means. We propose a perceptuo-analytic organization of sounds so that any environmental sound can be analyzed based on its signal characteristics and perception. We present environmental sound acquisition in the context of collection and annotation of a database for the footstep sounds, a common environmental sound, and show that it can be represented by these unconventional means and further analyzed to produce descriptions which are obscured with the traditional analysis. We present a novel application of extracting gait characteristics from the footstep sounds which is enabled by the proposed framework.

No description available.

Thesis (Ph. D.)--Department of Music, Stanford University, 2007. / Includes bibliographical references (leaves 226-236).

Traditional signal detection algorithms are often robust and are typically sufficient for high SNR data. However, the assumptions behind these methods begin to fall apart when signal period becomes either very short, or small in amplitude compared to any corruptive noise. To address this a kernel transform based cross-correlation algorithm is proposed for the application of weak signal detection. The algorithm leverages kernel methods to inflate SNR of the data and enhance the noise rejection capabilities of the traditional cross-correlation. The goal of the algorithm is to achieve detection for signals past the limits of those of the matched filter and the cross-correlation in the presence of white and colored noise. To evaluate the effectiveness of the correlation algorithm, Monte Carlo simulations are performed to determine the performance in the context of different types of noise. The performance of the algorithm will be compared against the cross-correlation and the matched filter. / Master of Science / As society advances, the tools we rely on become increasingly more intricate. Pivotal to the development of these systems is the algorithms used to process the data they collect. Particularly crucial to the field of signal processing, weak signal detection is focused on the processing of barely comprehensible data in the context of powerful noise. In recent years, advancements in weak signal detection have focused on pushing the theoretical limits of signal discernibility, especially when heavily obscured by noise. Leveraging the power of machine learning, certain AI algorithms have showcased promise in the detection of weak signals. It has yet to be seen if a foundational principle of AI called a kernel transform can be applied to classic signal detection theory to increase detection performance. This thesis will propose a kernel based detection algorithm for weak signal detection and the performance of the algorithm will be compared against previously established theory. New breakthroughs in detection algorithms facilitate improvements in active and passive sonar, medical devices and even the finance sector.

No description available.

No description available.

No description available.

No description available.

No description available.