Advanced Architectures for Signal Processors: (A Simple RNS (Residue Number System)/Systolic Architecture for Isolated Word Recognition)

摘要

Isolated work recognition is a computationally intensive processing function for which the combination of residue number system (RE4NS) computation and systolic array architecture offers practical simplification in the design of a special-purpose hardware processor. We have developed a simple architecture that uses short-time correlation analysis to form the spectral patterns, a distortion function employing squared Euclidean distance between the normalized correlation values of the test and reference utternace segments, and a two-dimensional pipelined processor array to implement dynamic time-wraping for pattern registration. With the exception of the normalization of the sample correlation values, all significant computations are carried out in a residue number system of compact size to be implemented in a specially designed hardware of low complexity. This combination of techniques provides for a very high processing throughput in simple hardware that can be used for real-time word recognition in a large vocabulry.