부 벡터 군집화를 통한 FPGA기반 음성 인식기의 수행 속도 향상

Abstract

Reducing the memory bandwidth is very critical for real-time implementation of a large vocabulary speech recognizer. In this paper, we implemented sub-vector clustering algorithm in hardware to alleviate the problem of much extreme memory access. The experimental result shows that this implementation accelerates acoustic likelihood computation twice faster, and increases the speed of overall recognition by 22%.