Thin client front-end processor for distributed speech recognition

Abstract

We present a front-end feature processor for distributed speech recognition for an integer-based DSP, and we employ block floating point and range reduction for the computation of elementary functions. We show that by reducing the numerical accuracy of the block floating point and the elementary functions, we are able to reduce the operational requirements to 12.6 wMOPs, 2. 4 kWords of RAM, 3.7 kWords of ROM. When used on a small vocabulary of 800 words 6.4 perplexity, and a large vocabulary of 20,200 words 102.5 perplexity, our optimized DSP front-end produces recognition accuracy comparable to an equivalent implementation on a floating point processor, without requiring a retrain of the recognition system with features produced by our DSP front-end.