A low memory bandwidth Gaussian mixture model (GMM) processor for 20,000-word real-time speech recognition FPGA system

Abstract

We propose a GMM processor for large vocabulary real-time continuous speech recognition. This processor achieves low operating frequency and low memory bandwidth using parallelization and vector look-ahead schemes, which are suitable to FPGA implementation. We designed the proposed processor on a Celoxica RC250 FPGA board, and confirmed that the required frequency and memory bandwidth for real-time operation are reduced by 89.8% and 84.2%, respectively. The 20,000-word real-time GMM computation is made at a frequency of 30.4 MHz and memory bandwidth of 47 Mbps, on the prototype.