A PARALLEL PIPELINE COMPUTER ARCHITECTURE FOR SPEECH PROCESSING.

摘要

The subject of this dissertation is a modular multicomputer architecture for digital speech (and signal) processing that is eminently suitable for implementation using Very Large Scale Integration (VLSI) techniques.; Processors built using this approach consist of several hardware identical computers (called Array Block Computers or ABC's) interconnected to form a number of functional block pipelines operating in parallel. Each ABC is a microprogrammable computer including a Central Processing Unit, Program Memory, Data Memory and a Communications Unit that handles communications with adjacent ABC's.; Data transfers between ABC's occur simultaneously for all ABC's at fixed predetermined intervals. All data exchanges are controlled totally by hardware and they are transparent to the programmer of the ABC's, greatly simplifying programming. The ABC's operate on whole frames of data performing operations such as windowing, the FFT, etc. The interconnecting topology for the ABC's is identical to the block diagram of the function performed. Performance increases linearly with the number of the ABC's in the system.; In order to demonstrate the feasibility of the proposed architecture, a prototype was built using the AMD 2901 4-bit slice microprocessor. Sufficient performance was obtained for complex real-time speech processing at a cost comparable to commercially available signal processors.; It is estimated that it will be possible to interface a whole ABC with performance superior to the current implementation on a silicon chip using 1 micron NMOS technology. A serial communication scheme is proposed that would allow such ABC's to be packaged in an 8-pin package. The very small required pinout and the fact that all ABC's are hardware identical (differing only in their programs) makes them very easy to mass produce using VLSI techniques. This will lead to very low cost, extremely high instruction execution rate computers, an essential requirement for most speech processing applications and especially speech recognition.