A reconfigurable co-processor for GMM-based classifier

This paper presents an efficient and reconfigurable co-processor to calculate Mahalanobis distance, which is the most computation-intensive part in the GMM (Gaussian Mixture Models)-based classifier. The Mahalanobis distance's calculation is divided into three parts (vector-vector subtraction, matrix-vector multiplication, and vector-vector multiplication) and these three parts can operate in a parallel way. The proposed architecture was implemented in Xilinx FPGA XC5VLX110T. Tested with a 358-state 3-mixture 39-feature 800-word HMM, co-processor operates at 35MHz to meet real-time requirement of speech recognition.