Sound

The recent uptick in senseless shootings in otherwise quiet and relatively safe environments is powerful evidence of the need, now more than ever, to reduce these occurrences. Artificial intelligence (AI) can play a significant role in deterring individuals from attempting these acts of violence. The installation of audio sensors can assist in the proper surveillance of surroundings linked to public safety, which is the first step toward AI-driven surveillance. With the increasing popularity of machine learning (ML) processes, systems are being developed and optimized to assist personnel in highly dangerous situations. In addition to saving innocent lives, supporting the capture of the responsible criminals is part of the AI algorithm that can be hosted in acoustic gunshot detection systems (AGDSs). Although there has been some speculation that these AGDSs produce a higher false positive rate (FPR) than reported in their specifications, optimizing the dataset used for the model’s training and testing will enhance its performance.
This dissertation proposes a new gunshot-like sound database that can be incorporated into a dataset for improved training and testing of a ML gunshot detection model. Reduction of the sample bias (that is, a bias in ML caused by an incomplete database) is achievable. The Mel frequency cepstral coefficient (MFCC) feature extraction process was utilized in this research. The uniform manifold and projection (UMAP) algorithm revealed that the MFCCs of this newly created database were the closest sounds to a gunshot sound, as compared to other gunshot-like sounds reported in literature. The UMAP algorithm reinforced the outcome derived from the calculation of the distances of the centroids of various gunshot-like sounds in MFCCs’ clusters. Further research was conducted into the feature reduction aspect of the gunshot detection ML model. Reducing a feature set to a minimum, while also maintaining a high accuracy rate, is a key parameter of a highly efficient model. Therefore, it is necessary for field deployed ML applications to be computationally light weight and highly efficient. Building on the discoveries of this research can lead to the development of highly efficient gunshot detection models.

The problem investigated in this thesis is that of an infinite, fluid-loaded, elastic cylindrical shell with an inhomogeneity of finite length excited by an acoustic plane wave. Seven inhomogeneities are considered to examine the parameters that influence the scattering. A full numerical approach and an iterative approach are developed to solve the shell and acoustic equations of motion expressed in the wavenumber domain. The response Green's function in the spatial domain is obtained using the hybrid analytical numerical technique, while the far-field scattered pressure is obtained by applying the Stationary Phase approximation. An analytical approach for the special case of a concentrated ring is developed, and the results compared to those from the full numerical solution. The range of applicability of the iterative approach is also investigated. The results show that the scattering pattern is a function of the spectral contents of the inhomogeneity distribution, and that the inhomogeneity mass influenced both the scattering pattern, and the scattering level. From the results it was also noted that an oblique angle of incidence steered the main lobe of the scattering pattern in the direction of the incoming acoustic wave. It is also demonstrated that the concentrated ring is usually a poor model to represent inhomogeneity of finite length.

This work explores the process of model-based classification of speech audio signals using low-level feature vectors. The process of extracting low-level features from audio signals is described along with a discussion of established techniques for training and testing mixture model-based classifiers and using these models in conjunction with feature selection algorithms to select optimal feature subsets. The results of a number of classification experiments using a publicly available speech database, the Berlin Database of Emotional Speech, are presented. This includes experiments in optimizing feature extraction parameters and comparing different feature selection results from over 700 candidate feature vectors for the tasks of classifying speaker gender, identity, and emotion. In the experiments, final classification accuracies of 99.5%, 98.0% and 79% were achieved for the gender, identity and emotion tasks respectively.