Date of Award
Spring 1978
Document Type
Dissertation - Restricted
Degree Name
Doctor of Philosophy (PhD)
Department
Biomedical Engineering
Abstract
The classification of any physical process requires some decoding of its characteristic features. Depending upon the complexity of both the process and the classification space, this feature extraction procedure constitutes a critical factor in the correct classification of the process. In this research, the speech signal is considered a physical process from which characteristic features may be determined and subsequently used for classification. In particular, three feature extraction mechanisms, known as linear prediction, the Fast Fourier Transform (FFT), and zero-crossing analysis, are evaluated and compared in an effort to determine their relative applicability for the automatic recognition of speech. All three of these methods have exhibited potential for automatic speech recognition. Various researchers, for various reasons, have selected one method or some combination of methods for use in their respective speech analysis systems. Results reported for these systems, dependent upon many factors, are understandably very diverse and occasionally even conflicting. Factors such as vocabulary, type and number of speakers, algorithmic implementation variations, varied test conditions, and a wide diversity of recognition schemes affect recognition results to such an extent that effective overall evaluation of the three feature extraction methods as single, critical system parameters is very difficult. No comprehensive comparative evaluation of linear prediction, FFT, and zero-crossing analysis methods has been published to date. Following the presentation of descriptive and theoretical formulations for each of the three feature extraction methods, and a variety of evaluative techniques, various algorithms belonging to each method are evaluated...