On techniques for the evaluation and simulation of parallel computer algorithms and architectures for speech understanding. (Volumes I-V)
Abstract
Speech understanding is extremely complex and requires extensive computation. The complexity precludes real-time operation on a conventional serial computer. The computational requirements can be met using parallel and distributed processing. This work addresses techniques to evaluate and simulate parallel algorithms and architectures for the design and development of a multiple processor real-time speech understanding system. A distributed parallel speech understanding architecture model is used. Processing is distributed among parallel machine knowledge source components. A technique to analyze data-independent parallel algorithms to obtain performance measures of their execution on a parallel processing system is presented. The critical path through a parallel algorithm is obtained and simulated. All major algorithm components can be analyzed. Simulation models, designed from detailed analyses of hardware specifications, are presented. The accuracy of the critical path simulation technique is verified by comparing simulation times with execution time measurements on the PASM parallel processing system prototype. Implementations of parallel fast Fourier transform programs were executed on PASM. Detailed execution time measurements were made for the programs and for program components. These measurements isolate hardware and software effects. Measurements on programs executing on a small number of processors were used to extrapolate accurately performance on a larger system. A stochastic model of English speech generates a stream of phonetically labeled frames with the same characteristics as real speech to be used for architecture design. Markov chain models are presented that accurately model the probability of occurrence, transition probabilities, and the mean and variance of the duration of phonemes and silence regions. Estimates of the number of frames necessary to realize the model's statistical distributions are included. Architecture designs of speech understanding system components are presented. The computational requirements of an acoustic processing architecture are determined from algorithm complexity. A labeling, segmentation, and lexical processing architecture is specified from a stochastic simulation of parallel algorithms using the English speech model as input. The techniques form a diverse, flexible, and accurate set of tools suitable for many areas of computer design. The techniques can be used to achieve significant insight into the operation of parallel algorithms and parallel computer architectures.
Degree
Ph.D.
Advisors
Jamieson, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.