A pipelined window implementation of multilayer networks with time-sequence input patterns for speech recognition

Neural networks (NNs), as processors of time-sequence patterns, have been successfully applied to several speaker-dependent speech recognition systems. They can be efficiently implemented by a pipelined architecture. In this paper, we explore its window structure as a computation component in the architecture. Two types of windows, parallel and serial data flow window, and their computation units with both feedforward and feedback paths are developed. The implementations of the windows are easily matched to the VLSI medium. Examples of applications to the NNs and performance analysis are given to illustrate effectiveness of the proposed window computation